Recent Publications

The first step in automated analysis of medical volumetric data is to detect slices, where specific body parts are located. In our project, we aimed to extract the pelvis regionfrom whole-body CT scans. Two deep learning approaches, namely, an unsupervised slice score regressor, and a supervised slice classification method, were evaluated on a relatively small-sized dataset. The result comparison showed that both methods could detect the region of interest with accuracy above 93{\%}. Although the straightforward classification method delivered more accurate results (accuracy of 99{\%}), sometimes it tended to output discontinuous regions, which can be solved by combination of both approaches.

The manufacturing industry is seeing an increase in demand for more custom-made, low-volume production. This type of production is rarely automated and is to a large extent still performed manually. To keep up with the competition and market demands, manufacturers will have to undertake the effort to automate such manufacturing processes. However, automating low-volume production is no small feat as the solution should be adaptable and future proof to unexpected changes in customers’ demands. In this paper, we propose a reconfigurable robot workcell aimed at automating low-volume production. The developed workcell can adapt to the changes in manufacturing processes by employing a number of passive, reconfigurable hardware elements, supported by the ROS-based, modular control software. To further facilitate and expedite the setup process, we integrated intuitive, user-friendly robot programming methods with the available hardware. The system was evaluated by implementing five production processes from different manufacturing industries.

In this paper, we employ a central pattern generator (CPG) driven radial basis function network (RBFN) based controller to learn optimized locomotion for a complex dung beetle-like robot using reinforcement learning approach called “Policy Improvement with Path Integrals (PI $$^2$$ )”. Our CPG driven RBFN controller is inspired by rhythmic dynamic movement primitives (DMPs). The controller can be also seen as an extension to a traditional CPG controller, which usually controls only the frequency of the motor patterns but not the shape. Our controller uses the CPG to control the frequency while the RBFN takes care of the shape of the motor patterns. In this paper, we only focus on the shape of the motor patterns and optimize those with respect to walking speed and energy efficiency. As a result, the robot can travel faster and consume less power than using only the CPG controller.

Setting up computer vision quality control tasks in a robot workcell is expensive, time consuming, and often requires expert knowledge. In this work, a highly adaptable approach to mitigate this issue is introduced. First, an ontology of atomic building blocks is defined, where each block represents one computer vision algorithm. Second, these blocks are used to generate a complex graph structure. The generation follows a set of rules, which can be shown to the inexperienced worker as a list of simple questions. Afterwards, this graph can be refined. Finally, the graph is integrated into the robotic assembly sequence, and the visual quality control procedure can be executed. Two industrial use cases show the feasibility of this approach. Thus, this work leverages computer vision aided quality control tasks in robot workcells.

Synaptic connections between neurons in the brain are dynamic because of continuously ongoing spine dynamics, axonal sprouting, and other processes. In fact, it was recently shown that the spontaneous synapse-autonomous component of spine dynamics is at least as large as the component that depends on the history of pre- and postsynaptic neural activity. These data are inconsistent with common models for network plasticity and raise the following questions: how can neural circuits maintain a stable computational function in spite of these continuously ongoing processes, and what could be functional uses of these ongoing processes? Here, we present a rigorous theoretical framework for these seemingly stochastic spine dynamics and rewiring processes in the context of reward-based learning tasks. We show that spontaneous synapse-autonomous processes, in combination with reward signals such as dopamine, can explain the capability of networks of neurons in the brain to configure themselves for specific computational tasks, and to compensate automatically for later changes in the network or task. Furthermore, we show theoretically and through computer simulations that stable computational performance is compatible with continuously ongoing synapse-autonomous changes. After reaching good computational performance it causes primarily a slow drift of network architecture and dynamics in task-irrelevant dimensions, as observed for neural activity in motor cortex and other areas. On the more abstract level of reinforcement learning the resulting model gives rise to an understanding of reward-driven network plasticity as continuous sampling of network configurations.

Even in todays world, where augmented reality glasses and 3d sensors become rapidly less expensive and widely more used, the most important sensor remains the 2d RGB camera. Every camera is an optical device and prone to sensor noise, especially in dark environments or environments with extreme high dynamic range. The here introduced filter removes a wide variation of noise, for example Gaussian noise and salt-and-pepper noise, but preserves edges. Due to the highly parallel structure of the method, the implementation on a GPU runs in real-time, allowing us to process standard images within tens of milliseconds. The filter is first tested on 2d image data and based on the Berkeley Image Dataset and Coco Dataset we outperform other standard methods. Afterwards, we show a generalization to arbitrary dimensions using noisy low level sensor data. As a result the filter can be used not only for image enhancement, but also for noise reduction on sensors like acceleremoters, gyroscopes, or GPS-trackers, which are widely used in robotic applications.

When looking at an object humans can quickly and efficiently assess which actions are possible given the scene context. This task remains hard for machines. Here we focus on manipulation actions and in the first part of this study define an object-action linked ontology for such context dependent affordance analysis. We break down every action into three hierarchical pre-condition layers starting on top with abstract object relations (which need to be fulfilled) and in three steps arriving at the movement primitives required to execute the action. This ontology will then, in the second part of this work, be linked to actual scenes. First the system looks at the scene and for any selected object suggests some actions. One will be chosen and, we use now a simple geometrical reasoning scheme by which this actions movement primitives will be filled with the specific parameter values, which are then executed by the robot. The viability of this approach will be demonstrated by analysing several scenes and a large number of manipulations.

Die Segmentierung verschiedener Strukturen im Korper ist eine der grundlegenden Operationen in der medizinischen Bildverarbeitung. In dieser Arbeit werden auf Machine Learning basierende Methoden zur Segmentierung medizinischer Bilder untersucht. Das Ziel ist es, in MRI-Scans die Trachea zu segmentieren. Jedoch soll in dieser Arbeit speziell die Effizienz der Algorithmen im Vordergrund stehen. Die verwendeten Ansatze basierten auf einer Deep Learning Architektur, welche zunachst individuell optimiert wird. Es konnte ein maximaler DICE-Koeffizient von (94.4 2.1)% erzielt werden. Zusatzlich kann festgestellt werden, dass die Segmentierung sehr effizient geschieht. Die Segmentierung von einmen Datensatz aus 40 Schichten dauert dabei weniger als eine Sekunde, wobei bei bisherigen Methoden es uber eine Minute benotigte.

Currently, flying robotic systems are in development for package delivery, aerial exploration in catastrophe areas, or maintenance tasks. While many flying robots are used in connection with powerful, stationary computing systems, the challenge in autonomous devices---especially in indoor-rescue or rural missions---lies in the need to do all processing internally on low power hardware. Furthermore, the device cannot rely on a well ordered or marked surrounding. These requirements make computer vision an important and challenging task for such systems. To cope with the cumulative problems of low frame rates in combination with high movement rates of the aerial device, a hyperbolic mirror is mounted on top of a quadrocopter, recording omnidirectional images, which can capture features during fast pose changes. The viability of this approach will be demonstrated by analysing several scenes. Here, we present a novel autonomous robot, which performs all computations online on low power embedded hardware and is therefore a truly autonomous robot. Furthermore, we introduce several novel algorithms, which have a low computational complexity and therefore enable us to refrain from external resources.

Human activity understanding has attracted much attention in recent years, because it plays a key role in a wide range of applications such as human-computer interfaces, visual surveillance, video indexing, intelligent humanoids robots, ambient intelligence and more. Activity understanding strongly benefits from fast, predictive action recognition. Here we present a new prediction algorithm for manipulation action classes in natural scenes. Manipulations are first represented by their temporal sequence of changing static and dynamic spatial relations between the objects that take part in the manipulation. This creates a transition matrix, called

Kurzfassung. Die Segmentierung von Hintergrund und Brustgewebe ist ein wichtiger Teil der Auswertung von Magnetresonanztomographie-Daten der Brust. Normalerweise wird diese von Arzten manuell durchgefuhrt. In dieser Arbeit wurde die Segmentierung hingegen mit einer U-net Architektur realisiert. Dabei wurden zwei Netzwerke trainiert und anschließend auf ein unbekanntes Testset, bestehend aus 8 Probandinnen, angewendet. Die so berechneten Segmentierungen wurden dann mit von Arzten manuell vorgenommenen verglichen. Das erste U-net nutzt keine weitere Vorverarbeitungsmethode und erreicht einen DSC von 0.91 0.09 (Mittelwert Standardabweichung). Beim zweiten Netzwerk wurde der N4ITK Bias Correction Algorithmus als Vorverarbeitungsmethode verwendet. Die Masken fur N4ITK konnen sehr grob sein und daher in einer spateren Anwendung von einem Arzt schnell erstellt werden. In dieser Konstellation wurde bei der Segmentierung des Testsets ein DSC von 0.98 0.05 erreicht. Die Segmentierungen benotigen daruber hinaus nach Anfertigung der Masken fur den Vorverarbeitungsalgorithmus 14s. Die Methode hat somit das Potential, Anwendung in der medizinischen Diagnostik zu finden.

Insects have various gaits with specific characteristics and can change their gaits smoothly in accordance with their speed. These gaits emerge from the embodied sensorimotor interactions that occur between the insects neural control and body dynamic systems through sensory feedback. Sensory feedback plays a critical role in coordinated movements such as locomotion, particularly in stick insects. While many previously developed insect models can generate different insect gaits, the functional role of embodied sensorimotor interactions in the interlimb coordination of insects remains unclear because of their complexity. In this study, we propose a simple physical model that is amenable to mathematical analysis to explain the functional role of these interactions clearly. We focus on a foot contact sensory feedback called phase resetting, which regulates leg retraction timing based on touchdown information. First, we used a hexapod robot to determine whether the distributed decoupled oscillators used for legs with the sensory feedback generate insect-like gaits through embodied sensorimotor interactions. The robot generated two different gaits and one had similar characteristics to insect gaits. Next, we proposed the simple model as a minimal model that allowed us to analyze and explain the gait mechanism through the embodied sensorimotor interactions. The simple model consists of a rigid body with massless springs acting as legs, where the legs are controlled using oscillator phases with phase resetting, and the governed equations are reduced such that they can be explained using only the oscillator phases with some approximations. This simplicity leads to analytical solutions for the hexapod gaits via perturbation analysis, despite the complexity of the embodied sensorimotor interactions. This is the first study to provide an analytical model for insect gaits under these interaction conditions. Our results clarified how this specific foot contact sensory feedback contributes to generation of insect-like ipsilateral interlimb coordination during hexapod locomotion.

Small and medium size enterprises (SMEs) often have small batch production. It leads to decreasing product lifetimes and also to more frequent product launches. In order to assist such production, a highly reconfigurable robot workcell is being developed. In this work, a visual inspection system designed for the robot workcell is presented and discussed in the context of the automotive light assembly example. The proposed framework is implemented using ROS and OpenCV libraries. We describe the hardware and software components of the framework and explain the systems benefits when compared to other commercial packages.

In large industrial greenhouses, plants are usually treated following well established protocols for watering, nutrients, and shading/light. While this is practical for the automation of the process, it does not tap the full potential for optimal plant treatment. To more efficiently grow plants, specific treatments according to the plant individual needs should be applied. Experienced human gardeners are very good at treating plants individually. Unfortunately, hiring a crew of gardeners to carry out this task in large greenhouses is not cost effective. In this work we present a cognitive system that integrates artificial intelligence (AI) techniques for decision-making with robotics techniques for sensing and acting to autonomously treat plants using a real-robot platform. Artificial intelligence techniques are used to decide the amount of water and nutrients each plant needs according to the history of the plant. Robotic techniques for sensing measure plant attributes (e.g. leaves) from visual information using 3D model representations. These attributes are used by the AI system to make decisions about the treatment to apply. Acting techniques execute robot movements to supply the plants with the specified amount of water and nutrients.

Recent studies have demonstrated that Drosophila melanogaster (briefly Drosophila) can successfully perform higher cognitive processes including second order olfactory conditioning. Understanding the neural mechanism of this behavior can help neuroscientists to unravel the principles of information processing in complex neural systems (e.g. the human brain) and to create efficient and robust robotic systems. In this work, we have developed a biologically-inspired spiking neural network which is able to execute both first and second order conditioning. Experimental studies demonstrated that volume signaling (e.g. by the gaseous transmitter nitric oxide) contributes to memory formation in vertebrates and invertebrates including insects. Based on the existing knowledge of odor encoding in Drosophila, the role of retrograde signaling in memory function, and the integration of synaptic and non-synaptic neural signaling, a neural system is implemented as Simulated fly. Simulated fly navigates in a two-dimensional environment in which it receives odors and electric shocks as sensory stimuli. The model suggests some experimental research on retrograde signaling to investigate neural mechanisms of conditioning in insects and other animals. Moreover, it illustrates a simple strategy to implement higher cognitive capabilities in machines including robots.

Walking animals produce adaptive interlimb coordination during locomotion in accordance with their situation. Interlimb coordination is generated through the dynamic interactions of the neural system, the musculoskeletal system, and the environment, although the underlying mechanisms remain unclear. Recently, investigations of the adaptation mechanisms of living beings have attracted attention, and bio-inspired control systems based on neurophysiological findings regarding sensorimotor interactions are being developed for legged robots. In this review, we introduce adaptive interlimb coordination for legged robots induced by various factors (locomotion speed, environmental situation, body properties, and task). In addition, we show characteristic properties of adaptive interlimb coordination, such as gait hysteresis and different time-scale adaptations. We also discuss the underlying mechanisms and control strategies to achieve adaptive interlimb coordination and the design principle for the control system of legged robots.

Obstructive sleep apnea (OSA) is a public health problem. Detailed analysis of the para-pharyngeal fat pads can help us to understand the pathogenesis of OSA and may mediate the intervention of this sleeping disorder. A reliable and automatic para-pharyngeal fat pads segmentation technique plays a vital role in investigating larger data bases to identify the anatomic risk factors for the OSA.

Abstract The continuous turnover of synapses observed in cortical networks poses a severe problem for storing long-term memory in the connectivity of these networks. Yet, in this chapter, we demonstrate that this can be resolved by considering connections consisting of multiple synapses. We show that, under certain conditions, the interaction of synaptic and structural plasticity induces a collective dynamics across all synapses between two neurons with stable states at zero and at multiple synapses. This dynamics leads to experimentally observed connectivity and enables long-term information storage despite synaptic turnover. Furthermore, the resulting connectivity can be controlled by external stimulations: very low or high stimulation levels quickly drive the neurons to become connected with zero or multiple synapses, respectively. Using this to actively store information on multisynaptic connections entails that information storage can be orders of magnitude faster than information retention under intermediate stimulation levels.

Fruit flies (Drosophila melanogaster) rely on their olfactory system to process environmental information. Although an extensive number of studies have revealed the basic molecular and cellular mechanisms underlying information processing as well as neural circuits of learning in the Drosophilas olfactory system, there are still many questions that are awaiting an answer. Specifically, linking behavior to underlying molecular mechanisms and neural circuitry are some of the challenges of modern neuroscience. In this review, we present some models which are based on available data of the Drosophila olfactory system to describe the role of physiological as well as structural parameters in information processing in the Drosophila olfactory system.

In this work we present a convolutional neural network-based (CNN) model that predicts future movements of a ball given a series of images depicting the ball and its environment. For training and evaluation, we use artificially generated images sequences. Two scenarios are analyzed: Prediction in a simple table tennis environment and a more challenging squash environment. Classical 2D convolution layers are compared with 3D convolution layers that extract the motion information of the ball from contiguous frames. Moreover, we investigate whether networks with stereo visual input perform better than those with monocular vision only. Our experiments suggest that CNNs can indeed predict physical behaviour with small error rates on unseen data but the performance drops for very complex underlying movements.

Dynamic neural fields (DNFs) are dynamical systems models that approximate the activity of large, homogeneous, and recurrently connected neural networks based on a mean field approach. Within dynamic field theory, the DNFs have been used as building blocks in architectures to model sensorimotor embedding of cognitive processes. Typically, the parameters of a DNF in an architecture are manually tuned in order to achieve a specific dynamic behavior (e.g. decision making, selection, or working memory) for a given input pattern. This manual parameters search requires expert knowledge and time to find and verify a suited set of parameters. The DNF parametrization may be particular challenging if the input distribution is not known in advance, e.g. when processing sensory information. In this paper, we propose the autonomous adaptation of the DNF resting level and gain by a learning mechanism of intrinsic plasticity (IP). To enable this adaptation, an input and output measure for the DNF are introduced, together with a hyper parameter to define the desired output distribution. The online adaptation by IP gives the possibility to pre-define the DNF output statistics without knowledge of the input distribution and thus, also to compensate for changes in it. The capabilities and limitations of this approach are evaluated in a number of experiments including a comparison of stochastic gradient with natural gradient descent.

The inclusion of robots in our society is imminent, such as service robots. Robots are now capable of reliably manipulating objects in our daily lives but only when combined with artificial intelligence (AI) techniques for planning and decision-making, which allow a machine to determine how a task can be completed successfully. To perform decision making, AI planning methods use a set of planning operators to code the state changes in the environment produced by a robotic action. Given a specific goal, the planner then searches for the best sequence of planning operators, i.e., the best plan that leads through the state space to satisfy the goal. In principle, planning operators can be hand-coded, but this is impractical for applications that involve many possible state transitions. An alternative is to learn them automatically from experience, which is most efficient when there is a human teacher. In this study, we propose a simple and efficient decision-making framework for this purpose. The robot executes its plan in a step-wise manner and any planning impasse produced by missing operators is resolved online by asking a human teacher for the next action to execute. Based on the observed state transitions, this approach rapidly generates the missing operators by evaluating the relevance of several cause-effect alternatives in parallel using a probability estimate, which compensates for the high uncertainty that is inherent when learning from a small number of samples. We evaluated the validity of our approach in simulated and real environments, where it was benchmarked against previous methods. Humans learn in the same incremental manner, so we consider that our approach may be a better alternative to existing learning paradigms, which require offline learning, a significant amount of previous knowledge, or a large number of samples.

Rhythmic neural signals serve as basis of many brain processes, in particular of locomotion control and generation of rhythmic movements. It has been found that specific neural circuits, named central pattern generators (CPGs), are able to autonomously produce such rhythmic activities. In order to tune, shape and coordinate the produced rhythmic activity, CPGs require sensory feedback, i.e., external signals. Nonlinear oscillators are a standard model of CPGs and are used in various robotic applications. A special class of nonlinear oscillators are adaptive frequency oscillators (AFOs). AFOs are able to adapt their frequency toward the frequency of an external periodic signal and to keep this learned frequency once the external signal vanishes. AFOs have been successfully used, for instance, for resonant tuning of robotic locomotion control. However, the choice of parameters for a standard AFO is characterized by a trade-off between the speed of the adaptation and its precision and, additionally, is strongly dependent on the range of frequencies the AFO is confronted with. As a result, AFOs are typically tuned such that they require a comparably long time for their adaptation. To overcome the problem, here, we improve the standard AFO by introducing a novel adaptation mechanism based on dynamical coupling strengths. The dynamical adaptation mechanism enhances both the speed and precision of the frequency adaptation. In contrast to standard AFOs, in this system, the interplay of dynamics on short and long time scales enables fast as well as precise adaptation of the oscillator for a wide range of frequencies. Amongst others, a very natural implementation of this mechanism is in terms of neural networks. The proposed system enables robotic applications which require fast retuning of locomotion control in order to react to environmental changes or conditions.

In this work, a framework for fully automated lung extraction from magnetic resonance imaging (MRI) inspiratory data that have been acquired within a on-going epidemiological child cohort study is presented. The methods main steps are intensity inhomogeneity correction, denoising, clustering, airway extraction and lung region refinement. The presented approach produces highly accurate results (Dice coefficients 95%), when compared to semi-automatically obtained masks, and has potential to be applied to the whole study data.

Abstract Trajectory generation methods play an important role in robotics since they are essential for the execution of actions. In this paper we present a novel trajectory generation method for generalization of accurate movements with boundary conditions. Our approach originates from optimal control theory and is based on a second order dynamic system. We evaluate our method and compare it to the state of the art movement generation methods in both simulations and real robot experiments. We show that the new method is very compact in its representation and can reproduce reference trajectories with zero error. Moreover, it has most of the features of the state of the art movement generation methods such as robustness to perturbations and generalization to new position and velocity boundary conditions. We believe that, due to these features, our method may have potential for robotic applications where high accuracy is required paired with flexibility, for example, in modern industrial robotic applications, where more flexibility will be demanded as well as in medical robotics.

Stroke is a leading cause of serious long-term disability. Post-stroke rehabilitation is a demanding task for the patient and a costly challenge for both society and healthcare systems. We present a novel approach for training of upper extremities after a stroke by utilising an industrial robotic arm and dynamic movement primitives (DMPs) with force feedback. We show how pre-recorded and learned DMPs can act as basis exercises, that can be modified into individualized and adaptive rehabilitation exercises that fit with the patients physical prop- erties and impairments. We conclude that our novel approach allows for easy and flexible set-up of rehabilitation exercises and has the potential to provide the therapists and patients much easier interaction with such complex technology.

The goal of this work is to densely predict a comparatively large set of affordances given only single RGB images. We approach this task by using a convolutional neural network based on the well-known ResNet architecture, which we blend with refinement modules recently proposed in the semantic segmentation literature. A novel cost function, capable of handling incomplete data, is introduced, which is necessary because we make use of segmentations of objects and their parts to generate affordance maps. We demonstrate both, quantitatively and qualitatively, that learning a dense predictor of affordances from an object part dataset is indeed possible and show that our model outperforms several baselines.

Function approximation in online, incremental, reinforcement learning needs to deal with two fundamental problems: biased sampling and non-stationarity. In this kind of tasks, biased sampling occurs because samples are obtained from specific trajectories dictated by the dynamics of the environment and are usually concentrated in particular convergence regions, which in the long term tend to dominate the approximation in the less sampled regions. The non-stationarity comes from the recursive nature of the estimations typical of temporal difference methods. This non-stationarity has a local profile, not only varying along the learning process but also along different regions of the state space. We propose to deal with these problems using an estimation of the probability density of samples represented with a Gaussian mixture model. To deal with the non-stationarity problem we use the common approach of introducing a forgetting factor in the updating formula. However, instead of using the same forgetting factor for the whole domain, we make it to depend on the local density of samples, which we use to estimate the non-stationarity of the function at any given input point. On the other hand, to address the biased sampling problem, the forgetting factor applied to each mixture component is modulated according to the new information provided in the updating, rather than forgetting only depending on time, thus avoiding undesired distortions of the approximation in less sampled regions.

During the last years, approaches based on convolutional neural networks (CNN) had substantial success in visual object perception. CNNs turned out to be capable of extracting high-level features of objects, which allow for fine-grained classification. However, some object classes exhibit tremendous variance with respect to their instances appearance. We believe that considering object parts as an intermediate representation could be helpful in these cases. In this work, a part-driven perception of everyday objects with a rotation estimation is implemented using deep convolution neural networks. The used network is trained and tested on artificially generated RGB-D data. The approach has a potential to be used for part recognition of realistic sensor recordings in present robot systems.

Recently, large population-based studies gain increasing focus in the research community. Epidemiological studies acquire numerous data by means of questionnaires and examinations. Many of these studies also collect imaging data, for instance, magnetic resonance imaging or ultrasonography from hundreds or even thousands of participants. Here, we consider several on-going epidemiological studies conducted in Europe as well as challenges of subsequent image analysis and visualization of heterogeneous data, which were obtained within these studies. In particular, the main focus is on airway extraction tasks and the visual analytics problems. Available solutions and future directions for computer science specialists are presented and analyzed in terms of user-friendliness, speed, and efficiency.

Recognition of human manipulation actions together with the analysis and execution by a robot is an important issue. Also, perception of spatial relationships between objects is central to understanding the meaning of manipulation actions. Here we would like to merge these two notions and analyze manipulation actions using symbolic spatial relations between objects in the scene. Specifically, we define procedures for extraction of symbolic human-readable relations based on Axis Aligned Bounding Box object models and use sequences of those relations for action recognition from image sequences. Our framework is inspired by the so called Semantic Event Chain framework, which analyzes touching and un-touching events of different objects during the manipulation. However, our framework uses fourteen spatial relations instead of two. We show that our relational framework is able to differentiate between more manipulation actions than the original Semantic Event Chains. We quantitatively evaluate the method on the MANIAC dataset containing 120 videos of eight different manipulation actions and obtain 97% classification accuracy which is 12 % more as compared to the original Semantic Event Chains.

Abstract The bursting of action potential and sparse activity are ubiquitously observed in the brain. Although the functions of these activity modes remain to be understood, it is expected that they play a critical role in information processing. In addition, the functional role of retrograde signalling in neural systems is under intensive research. Therefore, we propose a bio-inspired neural network that is capable of demonstrating these activity modes as well as shifting themselves from normal to bursting or sparse modes by changing model parameter values. Accordingly, we model diffused retrograde signalling with different activity patterns in dendrites and presynaptic neurons. Using in a three-layered spiking neural network, simulation studies are conducted using different conditions and parameter values to find factors underlying the change in firing rate of output neurons. Our findings propose the application of retrograde signalling as a known synaptic mechanism for the development of artificial neural systems to encode environmental information by different spiking modes.

We present a three-level cognitive system in a learning by demonstration context. The system allows for learning and transfer on the sensorimotor level as well as the planning level. The fundamentally different data structures associated with these two levels are connected by an efficient mid-level representation based on so-called semantic event chains. We describe details of the representations and quantify the effect of the associated learning procedures for each level under different amounts of noise. Moreover, we demonstrate the performance of the overall system by three demonstrations that have been performed at a project review. The described system has a technical readiness level (TRL) of 4, which in an ongoing follow-up project will be raised to TRL 6.

Working memory stores and processes information received as a stream of continuously incoming stimuli. This requires accurate sequencing and it remains puzzling how this can be reliably achieved by the neuronal system as our perceptual inputs show a high degree of temporal variability. One hypothesis is that accurate timing is achieved by purely transient neuronal dynamics by contrast a second hypothesis states that the underlying network dynamics are dominated by attractor states. In this study, we resolve this contradiction by theoretically investigating the performance of the system using stimuli with differently accurate timing. Interestingly, only the combination of attractor and transient dynamics enables the network to perform with a low error rate. Further analysis reveals that the transient dynamics of the system are used to process information, while the attractor states store it. The interaction between both types of dynamics yields experimentally testable predictions and we show that this way the system can reliably interact with a timing-unreliable Hebbian-network representing long-term memory. Thus, this study provides a potential solution to the long-standing problem of the basic neuronal dynamics underlying working memory.

A robot can feasibly be given knowledge of a set of tools for manipulation activities (e.g. hammer, knife, spatula). If the robot then operates outside a closed environment it is likely to face situations where the tool it knows is not available, but alternative unknown tools are present. We tackle the problem of finding the best substitute tool based solely on 3D vision data. Our approach handcodes simple models of known tools in terms of superquadrics and relationships among them. Our system attempts to fit these models to pointclouds of unknown tools, producing a numeric value for how good a fit is. This value can be used to rate candidate substitutes. We explicitly control how closely each part of a tool must match our model, under direction from parameters of a target task. We allow bottom-up information from segmentation to dictate the sizes that should be considered for various parts of the tool. These ideas allow for a flexible matching so that tools may be superficially quite different, but similar in the way that matters. We evaluate our systems ratings relative to other approaches and relative to human performance in the same task. This is an approach to knowledge transfer, via a suitable representation and reasoning engine, and we discuss how this could be extended to transfer in planning.

The control of multilegged animal walking is a neuromechanical process, and to achieve this in an adaptive and energy efficient way is a difficult and challenging problem. This is due to the fact that this process needs in real time: 1) to coordinate very many degrees of freedom of jointed legs 2) to generate the proper leg stiffness (i.e., compliance) and 3) to determine joint angles that give rise to particular positions at the endpoints of the legs. To tackle this problem for a robotic application, here we present a neuromechanical controller coupled with sensorimotor learning. The controller consists of a modular neural network for coordinating 18 joints and several virtual agonist-antagonist muscle mechanisms (VAAMs) for variable compliant joint motions. In addition, sensorimotor learning, including forward models and dual-rate learning processes, is introduced for predicting foot force feedback and for online tuning the VAAMs stiffness parameters. The control and learning mechanisms enable the hexapod robot advanced mobility sensor driven-walking device (AMOS) to achieve variable compliant walking that accommodates different gaits and surfaces. As a consequence, AMOS can perform more energy efficient walking, compared to other small legged robots. In addition, this paper also shows that the tight combination of neural control with tunable muscle-like functions, guided by sensory feedback and coupled with sensorimotor learning, is a way forward to better understand and solve adaptive coordination problems in multilegged locomotion.

In this paper, we address the problem of finding replacements of missing objects, involved in the execution of manipulation tasks. Our approach is based on estimating functional affordances for the unknown objects in order to propose replacements. We use a vision-based affordance estimation system utilizing object-wise global features and a multi-label learning method. This method also associates confidence values to the estimated affordances. We evaluate our approach on kitchen-related manipulation affordances. The evaluation also includes testing different scenarios for training the system using large-scale datasets. The results indicate that the system is able to successfully predict the affordances of novel objects. We also implement our system on a humanoid robot and demonstrate the affordance estimation in a real scene.

In this paper, we discuss magnetic resonance (MR) lung imaging and the related image processing tasks from two on-going epidemiological studies conducted in Germany. A modularized system for efficient lung segmentation is proposed and applied for test lung datasets from both studies. The efficiency of the framework is demonstrated by comparison of automatically computed results to the manually created ground truth masks. The presented pipeline allows one to obtain highly accurate segmentation results even for MR data with lower quality.

Magnetic resonance imaging (MRI) is a non-radiation based examination method, which gains an increasing popularity in research and clinical settings. Manual analysis of large data volumes is a very time-consuming and tedious process. Therefore, automatic analysis methods are required. This paper reviews different methods that have been recently proposed for automatic and semi-automatic lung segmentation from magnetic resonance imaging data. These techniques include thresholding, region growing, morphological operations, active contours, level sets, and neural networks. We also discuss the methodologies that have been utilized for performance and accuracy evaluation of each method.

Trajectory generation methods play an important role in robotics since they are essential for the execution of actions. In this paper we present a novel trajectory generation method for generalization of accurate movements with boundary conditions. Our approach originates from optimal control theory and is based on a second order dynamic system. We evaluate our method and compare it to state-of-the-art movement generation methods in both simulations and a real robot experiment. We show that the new method is very compact in its representation and can reproduce demonstrated trajectories with zero error. Moreover, it has most of the properties of the state-of-the-art trajectory generation methods such as robustness to perturbations and generalisation to new boundary position and velocity conditions. We believe that, due to these features, our method has great potential for various robotic applications, especially, where high accuracy is required, for example, in industrial and medical robotics.

Full-size humanoid robots are traditionally controlled with the Zero Moment Point (ZMP)-paradigm and simplified dynamics, a well established method which can be applied to balancing, walking, and whole-body manipulation tasks. For pure walking control, approaches like pattern generators and reflexes are employed, often on optimized hardware. Both controller groups are developed on different platforms and therefore can only be indirectly compared in terms of human likeness or energy efficiency. We present a reflex based neuro-controller with an underlying, simple hill-type muscle model on the extremely versatile humanoid robot ARMAR-4. We demonstrate the reflexive controllers flexible capabilities in terms of walking speed, step length, energy efficiency and inherent robustness against fall due to small slopes and pushes along the frontal axis. We contrast this controller with a Linearized Inverted Pendulum (LIP)-based ZMP-controller on the same platform. The promising results of this study show that even general humanoid robots can benefit from reflexive control schemes and encourage further investigation in this field.

In this paper a generic approach to the SLAM (Simultaneous Localization and Mapping) problem is proposed. The approach is based on a probabilistic SLAM algorithm and employs only two portable sensors, an inertial measurement unit (IMU) and a laser range finder (LRF) to estimate the state and environment of a robot. Scan-matching is applied to compensate for noisy IMU measurements. This approach does not require any robot-specific characteristics, e.g. wheel encoders or kinematic models. In principle, this minimal sensory setup can be mounted on different robot systems without major modifications to the underlying algorithms. The sensory setup with the probabilistic algorithm is tested in real-world experiments on two different kinds of robots: a simple two-wheeled robot and the six-legged hexapod AMOSII. The obtained results indicate a successful implementation of the approach and confirm its generic nature. On both robots, the SLAM problem can be solved with reasonable accuracy.

Animals show remarkable capabilities in navigating their habitat in a fully autonomous and energy-efficient way. In many species, these capabilities rely on a process called path integration, which enables them to estimate their current location and to find their way back home after long-distance journeys. Path integration is achieved by integrating compass and odometric cues. Here we introduce a neural path integration mechanism that interacts with a neural locomotion control to simulate homing behavior and path integration-related behaviors observed in animals. The mechanism is applied to a simulated six-legged artificial agent. Input signals from an allothetic compass and odometry are sustained through leaky neural integrator circuits, which are then used to compute the home vector by local excitation-global inhibition interactions. The home vector is computed and represented in circular arrays of neurons, where compass directions are population-coded and linear displacements are rate-coded. The mechanism allows for robust homing behavior in the presence of external sensory noise. The emergent behavior of the controlled agent does not only show a robust solution for the problem of autonomous agent navigation, but it also reproduces various aspects of animal navigation. Finally, we discuss how the proposed path integration mechanism may be used as a scaffold for spatial learning in terms of vector navigation.

We present an integrated system for the recognition, pose estimation and simultaneous tracking of multiple objects in 3D scenes. Our target application is a complete semantic representation of dynamic scenes which requires three essential steps recognition of objects, tracking their movements, and identification of interactions between them. We address this challenge with a complete system which uses object recognition and pose estimation to initiate object models and trajectories, a dynamic sequential octree structure to allow for full 6DOF tracking through occlusions, and a graph-based semantic representation to distil interactions. We evaluate the proposed method on real scenarios by comparing tracked outputs to ground truth trajectories and we compare the results to Iterative Closest Point and Particle Filter based trackers.

This study shows how understanding of object functionality arises by analyzing objects at the level of their parts where we focus here on primary tools. First, we create a set of primary tool functionalities, which we speculate is related to the possible functions of the human hand. The function of a tool is found by comparing it to this set. For this, the unknown tool is segmented, using a data-driven method, into its parts and evaluated using the geometrical part constellations against the training set. We demonstrate that various tools and even uncommon tool-versions can be recognized. The system

While humans can easily separate unknown objects into meaningful parts, recent segmentation methods can only achieve similar partitionings by training on human-annotated ground-truth data. Here we introduce a bottom-up method for segmenting 3D point clouds into functional parts which does not require supervision and achieves equally good results. Our method uses local concavities as an indicator for inter-part boundaries. We show that this criterion is efficient to compute and generalizes well across different object classes. The algorithm employs a novel locally constrained geometrical boundary model which proposes greedy cuts through a local concavity graph. Only planar cuts are considered and evaluated using a cost function, which rewards cuts orthogonal to concave edges. Additionally, a local clustering constraint is applied to ensure the partitioning only affects relevant locally concave regions. We evaluate our algorithm on recordings from an RGB-D camera as well as the Princeton Segmentation Benchmark, using a fixed set of parameters across all object classes. This stands in stark contrast to most reported results which require either knowing the number of parts or annotated ground-truth for learning. Our approach outperforms all existing bottom-up methods (reducing the gap to human performance by up to 50 %) and achieves scores similar to top-down data-driven approaches.

Walking animals, like stick insects, cockroaches or ants, demonstrate a fascinating range of locomotive abilities and complex behaviors. The locomotive behaviors can consist of a variety of walking patterns along with adaptation that allow the animals to deal with changes in environmental conditions, like uneven terrains, gaps, obstacles etc. Biological study has revealed that such complex behaviors are a result of a combination of biomechanics and neural mechanism thus representing the true nature of embodied interactions. While the biomechanics helps maintain flexibility and sustain a variety of movements, the neural mechanisms generate movements while making appropriate predictions crucial for achieving adaptation. Such predictions or planning ahead can be achieved by way of internal models that are grounded in the overall behavior of the animal. Inspired by these findings, we present here, an artificial bio-inspired walking system which effectively combines biomechanics (in terms of the body and leg structures) with the underlying neural mechanisms. The neural mechanisms consist of 1) central pattern generator based control for generating basic rhythmic patterns and coordinated movements, 2) distributed (at each leg) recurrent neural network based adaptive forward models with efference copies as internal models for sensory predictions and instantaneous state estimations, and 3) searching and elevation control for adapting the movement of an individual leg to deal with different environmental conditions. Using simulations we show that this bio-inspired approach with adaptive internal models allows the walking robot to perform complex locomotive behaviors as observed in insects, including walking on undulated terrains, crossing large gaps as well as climbing over high obstacles. Furthermore we demonstrate that the newly developed recurrent network based approach to sensorimotor prediction outperforms the previous state of the art adaptive neuron forward models.

A long-standing problem is how memories can be stored for very long times despite the volatility of the underlying neural substrate, most notably the high turnover of dendritic spines and synapses. To address this problem, here we are using a generic and simple probabilistic model for the creation and removal of synapses. We show that information can be stored for several months when utilizing the intrinsic dynamics of multi- synapse connections. In such systems, single synapses can still show high turnover, which enables fast learning of new information, but this will not perturb prior stored information (slow forgetting), which is represented by the compound state of the connections. The model matches the time course of recent experimental spine data during learning and memory in mice supporting the assumption of multi-synapse connections as the basis for long-term storage.

Motor primitives provide a modular organization to complex behaviours in both vertebrates and invertebrates. Inspired by this, here we generate motor primitives for a complex snake-like robot with screw-drive units, and thence chain and combine them, in order to provide a versatile, goal-directed locomotion for the robot. The behavioural primitives of the robot are generated using a reinforcement learning approach called

Abstract In this paper, we present a novel multi-level procedure for finding and tracking leaves of a rosette plant, in our case up to 3 weeks old tobacco plants, during early growth from infrared-image sequences. This allows measuring important plant parameters, e.g. leaf growth rates, in an automatic and non-invasive manner. The procedure consists of three main stages: preprocessing, leaf segmentation, and leaf tracking. Leaf-shape models are applied to improve leaf segmentation, and further used for measuring leaf sizes and handling occlusions. Leaves typically grow radially away from the stem, a property that is exploited in our method, reducing the dimensionality of the tracking task. We successfully tested the method on infrared image sequences showing the growth of tobacco-plant seedlings up to an age of about 30&#xa0days, which allows measuring relevant plant growth parameters such as leaf growth rate. By robustly fitting a suitably modified autocatalytic growth model to all growth curves from plants under the same treatment, average plant growth models could be derived. Future applications of the method include plant-growth monitoring for optimizing plant production in green houses or plant phenotyping for plant research.

Walking is so fundamental in everyday life that it is, for most people, an unconscious action. Loss or limitations in the ability to walk or stand directly impair our mobility and independence. Reasons of limitations can be stroke, paraplegia, or other damages to nerves, muscles, tendons, or limbs, encephalitis, brain abscesses, myopathies and further incidents and diseases affecting the motor control or the musculoskeletal system. In many cases, patients can be helped by, e.g., the use of orthoses for the lower limbs which assist to support the body and enable the patients to regain their movement abilities. Important factors and problems dominate the choice and usage of the suitable device: (i) Individualisation: The individual patients neurological status and remaining motor function have to be compatible with the support provided by the device. Particularly with regard to preserve---and not to interfere with---the remaining abilities, the device is selected to provide as little support as possible. As the remaining abilities largely vary with the individual expression of medical indications, the matching process is personalised and patient centred. (ii) Specialised Design: The movements a device supports are determined by its controller. Thereby, mobility is often limited to one or two basic movements, like walking and sitting. This specialisation imposes restrictions on the patients mobility. (iii) Target Group: The matching of the individuals need for assistance with the controllers abilities substantially restrict the target group of a device. (iv) Asymmetric Use: Patients often favour their healthy limb, leading to asymmetric gait and other gait deviations, implying consequential damage. (v) Device Acceptance and User Opinions: Device acceptance by its user is affected by many factors, as, for instance, comfort, the applicability in daily activities, cosmetic factors, and the patients impression if their opinions were considered in the process of device selection. Several studies indicate that, although a device might fit from an orthopaedic point of view, 60% up to nearly 100% of patients abandoned it for subjective reasons. Here, we assume that all these five problems can be addressed by the devices controller. So far, controllers are only used to tackle some of these problems isolated. We propose a modular controller architecture, which is designed for flexible use, expandability, and adaptation, e.g., learning from individually observed gait samples and intent recognition, solving the set of problems. The development was realised on a semi-active Knee-Ankle-Foot-Orthosis with hydraulic knee-damper and tested on a healthy walker. To address specialised design, we develop a controller based on a gait-independent formalism: An artificial neural network abstracts gait progress by decoding the sensory input. On top of this gait progress representation a device-specific network provides hardware control. To facilitate individualisation, the gait progress representation is learned from the patients gait samples, and a user interface allows direct user-interaction to define the control output, embedding the users opinions directly in the process to provide support for the individual motions. The use of artificial neural networks provides adaptation algorithms. The support of individual gaits leads itself to a specialisation of the controller. Here, we developed fast and reliable intent recognition with gait switching. The switching is done between per-gait modules, which consist of networks for gait progress abstraction, control output generation and internal models to predict gait dynamics. The prediction error identifies the optimal gait. This modular approach does not limit the number of movements, in contrast, it allows to extend the controller by further gaits in a formalised manner. It completes the solution to the problem of specialised design with a formalism which allows to extend the number of supported gaits with respect to the patients requirements. The proposed controller architecture focuses on the patients gait dynamics. The used sensors describe the joint dynamics and are not bound to a specific hardware-design. Tests on two variations of the presented orthosis prototype support this hypotheses. This reduces the requirements on the patients remaining abilities to the initiation of periodic motion with the support of the orthosis, expanding the target group. The support of individual gait allows the patients to develop their own gait, the patients do not have to force their gait into a pattern recognisable by the controller, providing a possibility for more symmetric gait. In a gait laboratory study, combining motion capture and electromyography, we investigated the user-device-interaction and how it alters the subjects gait. We found that 1. the deviations imposed by the hardware dominate those by the controller, 2. we located the upper body as the place with the largest deviations, and 3. we conclude that controller optimisation can be driven by a careful analysis of additional muscular activity in electromyographic recordings. This study shows that the presented controller supports the healthy walkers gait, but shows the limitations of the controllers impact due to hardware and sensory restrictions. The localisation of gait deviations identifies potential for manual and online controller-adaptation. To summarise, in this thesis we developed a controller on an orthosis prototype with a healthy walker based on a modular architecture allowing individual patient support. The system learns in a training process from observed gait samples and allows a simple and fast adaptation to gait changes and, in addition, enables easy extensions with further gaits. The evaluation of the user-device-interaction indicates deviations in the upper body and muscle work against the orthosis. This relation enables us in the next steps to infer how the devices support can be optimised and how an automatic adaptation mechanism can quantify its impact on the patients gait. Based on the here presented groundwork of an adaptive controller architecture, now it is possible to develop an observing, adapting controller, which is capable of basic patient surveillance, complementing medical treatment and rehabilitation.

In the online version of the EM algorithm introduced by Sato and Ishii (2000), a time-dependent discount factor is introduced for forgetting the effect of the old estimated values obtained with an earlier, inaccurate estimator. In their approach, forgetting is uniformly applied to the estimators of each mixture component depending exclusively on time, irrespective of the weight attributed to each unit for the observed sample. This causes an excessive forgetting in the less frequently sampled regions. To address this problem, we propose a modification of the algorithm that involves a weight-dependent forgetting, different for each mixture component, in which old observations are forgotten according to the actual weight of the new samples used to replace older values. A comparison of the time-dependent versus the weight-dependent approach shows that the latter improves the accuracy of the approximation and exhibits much greater stability.

Commonly complex cognitive concepts cannot consistently be connected to simple features of the world. Geometrical shape parameters and (e.g. edge features, compactness, color) may play a role for defining individual objects, but might be too variable to allow for concept formation. Earlier works had suggested that the formation of object concepts is strongly influenced by the division of our world along convex to concave surface transitions. In this first paper in a sequence of two we address this issue using abstract 3D geometrical structures (polycubes). In a first experiment, we let our subjects manipulate and compare polycubes with different compactness and different concavity/convexity asking which of them they would perceive as "an object". Both parameters (compactness and concavity/convexity) are not correlated in these stimuli. Nonetheless, we find that subjects with clear prevalence choose compact and convex ones. We continue to ask how strongly this influences the way we construct objects. Thus, in a second experiment we let humans combine polycubes to form an object. Also here we find that they prefer compact and convex configurations. This suggests that this simple geometric feature may underlie our cognitive understanding of object-ness not only with respect to perception but also by influencing how we build our world.

Objects usually consist of parts and the question arises whether there are perceptual features which allow breaking down an object into its fundamental parts without any additional (e.g. functional) information. As in the first paper of this sequence, we focus on the division of our world along convex to concave surface transitions. Here we are using machine vision to produce convex segments from 3D-scenes. We assume that a fundamental part is one, which we can easily name while at the same time there is no natural subdivision possible into smaller parts. Hence in this experiment we presented the computer vision generated segments to our participants and asked whether they can identify and name them. Additionally we control against segmentation reliability and we find a clear trend that reliable convex segments have a high degree of name-ability. In addition, we observed that using other image-segmentation methods will not yield nameable entities. This indicates that convex-concave surface transition may indeed form the basis for dividing objects into meaningful entities. It appears that other or further subdivisions do not carry such a strong semantical link to our everyday language as there are no names for them.

Abstract In this paper we present a hierarchical framework for representation of manipulation actions and its applicability to the problem of top down action extraction from observation. The framework consists of novel probabilistic semantic models, which encode contact relations as probability distributions over the action phase. The models are action descriptive and can be used to provide probabilistic similarity scores for newly observed action sequences. The lower level of the representation consists of parametric hidden Markov models, which encode trajectory information.

Human manipulation activity recognition is an important yet challenging task in robot imitation. In this paper, we introduce, for the first time, a novel method for semantic decomposition and recognition of continuous human manipulation activities by using on-line learned individual manipulation models. Solely based on the spatiotemporal interactions between objects and hands in the scene, the proposed framework can parse not only sequential and concurrent (overlapping) manipulation streams but also basic primitive elements of each detected manipulation. Without requiring any prior object knowledge, the framework can furthermore extract object-like scene entities that are performing the same role in the detected manipulations. The framework was evaluated on our new egocentric activity dataset which contains 120 different samples of 8 single atomic manipulations (e.g. Cutting and Stirring) and 20 long and complex activity demonstrations such as

In this work we address the problem of indoor scene understanding from RGB-D images. Specifically, we propose to find instances of common furniture classes, their spatial extent, and their pose with respect to generalized class models. To accomplish this, we use a deep, wide, multi-output convolutional neural network (CNN) that predicts class, pose, and location of possible objects simultaneously. To overcome the lack of large annotated RGB-D training sets (especially those with pose), we use an on-the-fly rendering pipeline that generates realistic cluttered room scenes in parallel to training. We then perform transfer learning on the relatively small amount of publicly available annotated RGB-D data, and find that our model is able to successfully annotate even highly challenging real scenes. Importantly, our trained network is able to understand noisy and sparse observations of highly cluttered scenes with a remarkable degree of accuracy, inferring class and pose from a very limited set of cues. Additionally, our neural network is only moderately deep and computes class, pose and position in tandem, so the overall run-time is significantly faster than existing methods, estimating all output parameters simultaneously in parallel on a GPU in seconds.

Robotic applications in human environments are usually implemented using a cognitive architecture that integrates techniques of different levels of abstraction, ranging from artificial intelligence techniques for making decisions at a symbolic level to robotic techniques for grounding symbolic actions. In this work we address the problem of simultaneous learning at different levels of abstractions in such an architecture. This problem is important since human environments are highly variable, and many unexpected situations may arise during the execution of a task. The usual approach under this circumstance is to train each level individually to learn how to deal with the new situations. However, this approach is limited since it implies long task interruptions every time a new situation needs to be learned. We propose an architecture where learning takes place simultaneously at all the levels of abstraction. To achieve this, we devise a method that permits higher levels to guide the learning at the levels below for the correct execution of the task. The architecture is instantiated with a logic-based planner and an online planning operator learner, at the highest level, and with online reinforcement learning units that learn action policies for the grounding of the symbolic actions, at the lowest one. A human teacher is involved in the decision-making loop to facilitate learning. The framework is tested in a physically realistic simulation of the Sokoban game.

In this paper we propose a novel spatially stratified sampling technique for evaluating the likelihood function in particle filters. In particular, we show that in the case where the measurement function uses spatial correspondence, we can greatly reduce computational cost by exploiting spatial structure to avoid redundant computations. We present results which quantitatively show that the technique permits equivalent, and in some cases, greater accuracy, as a reference point cloud particle filter at significantly faster run-times. We also compare to a GPU implementation, and show that we can exceed their performance on the CPU. In addition, we present results on a multi-target tracking appli- cation, demonstrating that the increases in efficiency permit online 6DoF multi-target tracking on standard hardware.

Walking animals, like insects, with little neural computing can effectively perform complex behaviors. They can walk around their environment, escape from corners/deadlocks, and avoid or climb over obstacles. While performing all these behaviors, they can also adapt their movements to deal with an unknown situation. As a consequence, they successfully navigate through their complex environment. The versatile and adaptive abilities are the result of an integration of several ingredients embedded in their sensorimotor loop. Biological studies reveal that the ingredients include neural dynamics, plasticity, sensory feedback, and biomechanics. Generating such versatile and adaptive behaviors for a walking robot is a challenging task. In this study, we present a bio-inspired approach to solve this task. Specifically, the approach combines neural mechanisms with plasticity, sensory feedback, and biomechanics. The neural mechanisms consist of adaptive neural sensory processing and modular neural locomotion control. The sensory processing is based on a small recurrent network consisting of two fully connected neurons. Online correlation-based learning with synaptic scaling is applied to adequately change the connections of the network. By doing so, we can effectively exploit neural dynamics (i.e., hysteresis effects and single attractors) in the network to generate different turning angles with short-term memory for a biomechanical walking robot. The turning information is transmitted as descending steering signals to the locomotion control which translates the signals into motor actions. As a result, the robot can walk around and adapt its turning angle for avoiding obstacles in different situations as well as escaping from sharp corners or deadlocks. Using backbone joint control embedded in the locomotion control allows the robot to climb over small obstacles. Consequently, it can successfully explore and navigate in complex environments.

The ability to quantify temporal information on the scale of hundreds of milliseconds is critical towards the processing of complex sensory and motor patterns. However, the nature of neural mechanisms for temporal information processing (at this scale) in the brain still remains largely unknown. Furthermore, given that biological organisms are situated in a dynamic environment, the processing of time-varying environmental stimuli is intricately related to the generation of cognitive behaviors, and as such, an important element of learning and memory. In order to model such temporal processing recurrent neural networks emerge as natural candidates due to their inherent dynamics and fading memory of advent stimuli. As such, this thesis investigates recurrent neural network (RNN) models driven by external stimuli as the basis of time perception and temporal processing in the brain. Such processing lies in the short timescale that is responsible for the generation of short-term memory-guided behaviors like complex motor pattern processing and generation, motor prediction, time-delayed responses, and goal-directed decision making. We present a novel self-adaptive RNN model and verify its ability to generate such complex temporally dependent behaviors, juxtaposing it critically with current state of the art non-adaptive or static RNN models. Taking into consideration the brains ability to undergo changes at structural and functional levels across a wide range of time spans, in this thesis, we make the primary hypothesis, that a combination of neuronal plasticity and homeostatic mechanisms in conjunction with the innate recurrent loops in the underlying neural circuitry gives rise to such temporally-guided actions. Furthermore, unlike most previous studies of spatio-temporal processing in the brain, here we follow a closed-loop approach. Such that, there is a tight coupling between the neural computations and the resultant behaviors, demonstrated on artificial robotic agents as the embodied self of a biological organism. In the first part of the thesis, using a RNN model of rate-coded neurons starting with random initialization of synaptic connections, we propose a learning rule based on local active information storage (LAIS). This is measured at each spatiotemporal location of the network, and used to adapt the individual neuronal decay rates or time constants with respect to the incoming stimuli. This allows an adaptive timescale of the network according to changes in timescales of inputs. We combine this, with a mathematically derived, generalized mutual information driven intrinsic plasticity mechanism that can tune the non-linearity of network neurons. This enables the network to maintain homeostasis as well as, maximize the flow of information from input stimuli to neuronal outputs. These unsupervised local adaptations are then combined with supervised synaptic plasticity in order to tune the otherwise fixed synaptic connections, in a task dependent manner. The resultant plastic network, significantly outperforms previous static models for complex temporal processing tasks in non-linear computing power, temporal memory capacity, noise robustness as well as tuning towards near-critical dynamics. These are displayed using a number of benchmark tests, delayed memory guided responses with a robotic agent in real environment and complex motor pattern generation tasks. Furthermore, we also demonstrate the ability of our adaptive network to generate clock like behaviors underlying time perception in the brain. The model output matches the linear relationship of variance and squared time interval as observed from experimental studies. In the second part of the thesis, we first demonstrate the application of our model on behaviorally relevant motor prediction tasks with a walking robot, implementing distributed internal forward models using our adaptive network. Following this, we extend the previous supervised learning scheme, by implementing reward-based learning following the temporal-difference paradigm, in order to adapt the synaptic connections in our network. The neuronal correlates of this formulation is discussed from the point of view of the cortico-striatal circuitry, and a new combined learning rule is presented. This leads to novel results demonstrating how the striatal circuitry works in combination with the cerebellar circuitry in the brain, that lead to robust goal-directed behaviors. Thus, we demonstrate the application of our adaptive network model on the entire spectrum of temporal information processing, in the timescale of few hundred milliseconds (complex motor processing) to minutes (delayed memory and decision making). Overall, the results obtained in this thesis affirms our primary hypothesis that plasticity and adaptation in recurrent networks allow complex temporal information processing, which otherwise cannot be obtained with purely static networks. Furthermore, homeostatic plasticity and neuronal timescale adaptations could be potential mechanisms by which the brain performs such processing with remarkable ease.

titleAuthor Summary/title pThe connectivity between neurons is modified by different mechanisms. On a time scale of minutes to hours one finds synaptic plasticity, whereas mechanisms for structural changes at axons or dendrites may take days. One main factor determining structural changes is the weight of a connection, which, in turn, is adapted by synaptic plasticity. Both mechanisms, synaptic and structural plasticity, are influenced and determined by the activity pattern in the network. Hence, it is important to understand how activity and the different plasticity mechanisms influence each other. Especially how activity influences rewiring in adult networks is still an open question./p pWe present a model, which captures these complex interactions by abstracting structural plasticity with weight-dependent probabilities. This allows for calculating the distribution of the number of synapses between two neurons analytically. We report that biologically realistic connection patterns for different cortical layers generically arise with synaptic plasticity rules in which the synaptic weights grow with postsynaptic activity. The connectivity patterns also lead to different activity levels resembling those found in the different cortical layers. Interestingly such a system exhibits a hysteresis by which connections remain stable longer than expected, which may add to the stability of information storage in the network./p

Cortical connectivity emerges from the permanent interaction between neuronal activity and synaptic as well as structural plasticity. An important experimentally observed feature of this connectivity is the distribution of the number of synapses from one neuron to another, which has been measured in several cortical layers. All of these distributions are bimodal with one peak at zero and a second one at a small number (3-8) of synapses. In this study, using a probabilistic model of structural plasticity, which depends on the synaptic weights, we explore how these distributions can emerge and which functional consequences they have. We find that bimodal distributions arise generically from the interaction of structural plasticity with synaptic plasticity rules that fulfill the following biological realistic constraints: First, the synaptic weights have to grow with the postsynaptic activity. Second, this growth curve and/or the input-output relation of the postsynaptic neuron have to change sub-linearly (negative curvature). As most neurons show such input-output-relations, these constraints can be fulfilled by many biological reasonable systems. Given such a system, we show that the different activities, which can explain the layer-specific distributions, correspond to experimentally observed activities. Considering these activities as working point of the system and varying the pre- or postsynaptic stimulation reveals a hysteresis in the number of synapses. As a consequence of this, the connectivity between two neurons can be controlled by activity but is also safeguarded against overly fast changes. These results indicate that the complex dynamics between activity and plasticity will, already between a pair of neurons, induce a variety of possible stable synaptic distributions, which could support memory mechanisms.

When learning a complex task our nervous system self-organizes large groups of neurons into coherent dynamic activity patterns. During this, a network with multiple, simultaneously active, and computationally powerful cell assemblies is created. How such ordered structures are formed while preserving a rich diversity of neural dynamics needed for computation is still unknown. Here we show that the combination of synaptic plasticity with the slower process of synaptic scaling achieves (i) the formation of cell assemblies and (ii) enhances the diversity of neural dynamics facilitating the learning of complex calculations. Due to synaptic scaling the dynamics of different cell assemblies do not interfere with each other. As a consequence, this type of self-organization allows executing a difficult, six degrees of freedom, manipulation task with a robot where assemblies need to learn computing complex non-linear transforms and - for execution - must cooperate with each other without interference. This mechanism, thus, permits the self-organization of computationally powerful sub-structures in dynamic networks for behavior control.

Image based object classification requires clean training data sets. Gathering such sets is usually done manually by humans, which is time-consuming and laborious. On the other hand, directly using images from search engines creates very noisy data due to ambiguous noun-focused indexing. However, in daily speech nouns and verbs are always coupled. We use this for the automatic generation of clean data sets by the here-presented TRANSCLEAN algorithm, which through the use of multiple languages also solves the problem of polysemes (a single spelling with multiple meanings). Thus, we use the implicit knowledge contained in verbs, e.g. in an imperative such as

In this work we address the problem of finding replacements of missing objects that are needed for the execution of human-like manipulation tasks. This is a usual problem that is easily solved by humans provided their natural knowledge to find object substitutions: using a knife as a screwdriver or a book as a cutting board. On the other hand, in robotic applications, objects required in the task should be included in advance in the problem definition. If any of these objects is missing from the scenario, the conventional approach is to manually redefine the problem according to the available objects in the scene. In this work we propose an automatic way of finding object substitutions for the execution of manipulation tasks. The approach uses a logic-based planner to generate a plan from a prototypical problem definition and searches for replacements in the scene when some of the objects involved in the plan are missing. This is done by means of a repository of objects and attributes with roles, which is used to identify the affordances of the unknown objects in the scene. Planning actions are grounded using a novel approach that encodes the semantic structure of manipulation actions. The system was evaluated in a KUKA arm platform for the task of preparing a salad with successful results.

The development of programming paradigms for industrial assembly currently gets fresh impetus from approaches in human demonstration and programming-by-demonstration. Major low- and mid-level prerequisites for machine vision and learning in these intelligent robotic applications are pose estimation, stereo reconstruction and action recognition. As a basis for the machine vision and learning involved, pose estimation is used for deriving object positions and orientations and thus target frames for robot execution. Our contribution introduces and applies a novel benchmark for typical multi-sensor setups and algorithms in the field of demonstration-based automated assembly. The benchmark platform is equipped with a multi-sensor setup consisting of stereo cameras and depth scanning devices (see Fig. 1). The dimensions and abilities of the platform have been chosen in order to reflect typical manual assembly tasks. Following the eRobotics methodology, a simulatable 3D representation of this platform was modelled in virtual reality. Based on a detailed camera and sensor simulation, we generated a set of benchmark images and point clouds with controlled levels of noise as well as ground truth data such as object positions and time stamps. We demonstrate the application of the benchmark to evaluate our latest developments in pose estimation, stereo reconstruction and action recognition and publish the benchmark data for objective comparison of sensor setups and algorithms in industry.

The goal-directed navigational ability of animals is an essential prerequisite for them to survive. They can learn to navigate to a distal goal in a complex environment. During this long-distance navigation, they exploit environmental features, like landmarks, to guide them towards their goal. Inspired by this, we develop an adaptive landmark-based navigation system based on sequential reinforcement learning. In addition, correlation-based learning is also integrated into the system to improve learning performance. The proposed system has been applied to simulated simple wheeled and more complex hexapod robots. As a result, it allows the robots to successfully learn to navigate to distal goals in complex environments.

Object recognition plays an important role in robotics, since objects/tools first have to be identified in the scene before they can be manipulated/used. The performance of object recognition largely depends on the training dataset. Usually such training sets are gathered manually by a human operator, a tedious procedure, which ultimately limits the size of the dataset. One reason for manual selection of samples is that results returned by search engines often contain irrelevant images, mainly due to the problem of homographs (words spelled the same but with different meanings). In this paper we present an automated and unsupervised method, coined Trainingset Cleaning by Translation ( TCT ), for generation of training sets which are able to deal with the problem of homographs. For disambiguation, it uses the context provided by a command like "tighten the nut" together with a combination of public image searches, text searches and translation services. We compare our approach against plain Google image search qualitatively as well as in a classification task and demonstrate that our method indeed leads to a task-relevant training set, which results in an improvement of 24.1% in object recognition for 12 ambiguous classes. In addition, we present an application of our method to a real robot scenario.

Neurobiological studies have shown that insects are able to adapt leg movements and posture for obstacle negotiation in changing environments. Moreover, the distance to an obstacle where an insect begins to climb is found to be a major parameter for successful obstacle negotiation. Inspired by these findings, we present an adaptive neural control mechanism for obstacle negotiation behavior in hexapod robots. It combines locomotion control, backbone joint control, local leg reflexes, and neural learning. While the first three components generate locomotion including walking and climbing, the neural learning mechanism allows the robot to adapt its behavior for obstacle negotiation with respect to changing conditions, e.g., variable obstacle heights and different walking gaits. By successfully learning the association of an early, predictive signal conditioned stimulus, CS and a late, reflex signal unconditioned stimulus, UCS, both provided by ultrasonic sensors at the front of the robot, the robot can autonomously find an appropriate distance from an obstacle to initiate climbing. The adaptive neural control was developed and tested first on a physical robot simulation, and was then successfully transferred to a real hexapod robot, called AMOS II. The results show that the robot can efficiently negotiate obstacles with a height up to 85% of the robots leg length in simulation and 75% in a real environment

Cognitive ageing seems to be a story of global degradation. As one ages there are a number of physical, chemical and biological changes that take place. Therefore it is logical to assume that the brain is no exception to this phenomenon. The principle purpose of this project is to use models of neural dynamics and learning based on the underlying principle of self-organised criticality, to account for the age related cognitive effects. In this regard learning in neural networks can serve as a model for the acquisition of skills and knowledge in early development stages i.e. the ageing process and criticality in the network serves as the optimum state of cognitive abilities. Possible candidate mechanisms for ageing in a neural network are loss of connectivity and neurons, increase in the level of noise, reduction in white matter or more interestingly longer learning history and the competition among several optimization objectives. In this paper we are primarily interested in the affect of the longer learning history on memory and thus the optimality in the brain. Hence it is hypothesized that prolonged learning in the form of associative memory patterns can destroy the state of criticality in the network. We base our model on Tsodyks and Markrams 49 model of dynamic synapses, in the process to explore the effect of combining standard Hebbian learning with the phenomenon of Self-organised criticality. The project mainly consists of evaluations and simulations of networks of integrate and fire-neurons that have been subjected to various combinations of neural-level ageing effects, with the aim of establishing the primary hypothesis and understanding the decline of cognitive abilities due to ageing, using one of its important characteristics, a longer learning history.1

In this paper we show that exchanging curved feet and rigid ankles by at feet and compliant ankles improves the range of gait parameters for a bipedal dynamic walker. The new lower legs were designed such that they t to the old set-up, allowing for a direct and quantitative comparison. The dynamic walking robot RunBot, controlled by an re exive neural network, uses only few sensors for generating its stable gait. The results show that at feet and compliant ankles extend RunBots parameter range especially to more leaning back postures. They also allow the robot to stably walk over obstacles with low height.

The idea that connected convex surfaces, separated by concave boundaries, play an important role for the perception of objects and their decomposition into parts has been discussed for a long time. Based on this idea, we present a new bottom-up approach for the segmentation of 3D point clouds into object parts. The algorithm approximates a scene using an adjacency-graph of spatially connected surface patches. Edges in the graph are then classified as either convex or concave using a novel, strictly local criterion. Region growing is employed to identify locally convex connected subgraphs, which represent the object parts. We show quantitatively that our algorithm, although conceptually easy to graph and fast to compute, produces results that are comparable to far more complex state-of-the-art methods which use classification, learning and model fitting. This suggests that convexity/concavity is a powerful feature for object partitioning using 3D data. Furthermore we demonstrate that for many objects a natural decomposition into

Robots are expected to operate autonomously in unconstrained, real-world environments. Therefore, they cannot rely on access to models of all objects in their environment, in order to parameterize object-directed actions. The robot must estimate the shape of objects in such environments, based on their perception. How to estimate an objects shape based on distal sensors, such as color- or depth cameras, has been extensively studied. Using haptic sensors for this purpose, however, has not been considered in a comparable depth. Humans, to the contrary, are able to improve object manipulation capabilities by using tactile stimuli, acquired from an active haptic exploration of an object. In this paper we introduce a neural-dynamic model which allows to build an object shape representation based on haptic exploration. Acquiring this representation during object manipulation requires the robot to autonomously detect and correct errors in the localization of tactile features with respect to the object. We have implemented an architecture for haptic exploration of an objects shape on a physical robotic hand in a simple exemplary scenario, in which the geometrical models of two different n-gons are learned from tactile data while rotating them with the robotic hand.

There have been several contributions on human motion detection and action recognition over the past two decades. However detection of abnormal and suspicious behaviors in video surveillance is currently one of the most interesting studies for many research groups in computer vision and artificial intelligence. There are two well-known models to detect suspicious behaviors misuse detection model and anomaly detection model. Misuse detection model is related to definition of suspicious behavior while anomaly detection model measures the difference between the defined normal behaviors and the current behavior. We employed the first model to classify human behaviors into normal, abnormal and suspicious types according to trajectory and spatio-temporal domains. In the former we define some abnormal trajectories, like crinkle or loitering trajectories and then compare patterns of input trajectories for each person with the predefined abnormal trajectory. In the second domain some special regions on a scene which are not usual for walking are predefined. If a person stays there more than a threshold time, his behavior will be assumed as an abnormal one. Under some conditions abnormal type will be interpreted as a suspicious type. This is done by introducing a "normality level" which determines the commonality level of each behavior. If a behavior has a high "normality level" value, it is normal and lower values show abnormal and suspicious types of behavior. Employing both domains simultaneously provides high degree of accuracy in the proposed approach. In addition, introducing several level of abnormality according to the "normality level" parameter and having fuzzy approach led to differentiate between warning and alarm states. The above points plus on-line working of the system opposite of the complexity of its algorithms are positive points of our work. Our method detects abnormal and suspicious behaviors in the CAVIAR data set with an accuracy of 90% in real time.

Today most recognition pipelines are trained at an off-line stage, providing systems with pre-segmented images and predefined objects, or at an on-line stage, which requires a human supervisor to tediously control the learning. Self-Supervised on-line training of recognition pipelines without human intervention is a highly desirable goal, as it allows systems to learn unknown, environment specific objects on-the-fly. We propose a fast and automatic system, which can extract and learn unknown objects with minimal human intervention by employing a two-level pipeline combining the advantages of RGB-D sensors for object extraction and high-resolution cameras for object recognition. Furthermore, we significantly improve recognition results with local features by implementing a novel keypoint orientation scheme, which leads to highly invariant but discriminative object signatures. Using only one image per object for training, our system is able to achieve a recognition rate of 79% for 18 objects, benchmarked on 42 scenes with random poses, scales and occlusion, while only taking 7 seconds for the training. Additionally, we evaluate our orientation scheme on the state-of-the-art 56-object SDU-dataset boosting accuracy for one training view per object by +37% to 78% and peaking at a performance of 98% for 11 training views.

Goal-directed decision making in biological systems is broadly based on associations between conditional and unconditional stimuli. This can be further classified as classical conditioning (correlation based learning) and operand conditioning (reward-based learning). A number of computational and experimental studies have well established the role of the basal ganglia (striatal system) towards reward-based learning, where as the cerebellum evidently plays an important role in developing specific conditioned responses. Although, they are viewed as distinct learning systems 1, recent animal experiments point towards their complementary role in behavioral learning, and also show the existence of substantial two-way communication between the two structures 2. Based on this notion of co-operative learning, in this work we hypothesize that the basal ganglia and cerebellar learning systems work in parallel and compete with each other (Figure 1). We envision such an interaction being driven by a simple reward modulated heterosynaptic plasticity (RMHP) rule 3, in order to guide the over all goal-directed behavior. Using a recurrent neural network actor-critic model of the basal ganglia and feed-forward correlation learning model of the cerebellum (input correlation learning-ICO) 4, we demonstrate that the RMHP rule can effectively combine the outcomes of the two learning systems. This is tested using simulated environments of increasing complexity with a four-wheeled animat in a dynamic foraging task. Although, they are modeled within a highly simplified level of biological abstraction, we clearly demonstrate that such a combined learning mechanism, leads to much stabler and faster learning of goal-directed behaviors in comparison to the individual systems.

Patients use orthoses and prosthesis for the lower limbs to support and enable movements, they can not or only with difficulties perform themselves. Because traditional devices support only a limited set of movements, patients are restricted in their mobility. A possible approach to overcome such limitations is to supply the patient via the orthosis with situation-dependent gait models. To achieve this, we present a method for gait recognition using model invalidation. We show that these models are capable to predict the individual patients movements and supply the correct gait. We investigate the systems accuracy and robustness on a Knee-Ankle-Foot-Orthosis, introducing behaviour changes depending on the patients current walking situation. We conclude that the here presented model-based support of different gaits has the power to enhance the patients mobility.

We propose a cognitive system that combines artificial intelligence techniques for planning and learning to execute tasks involving delayed and variable correlations between the actions executed and their expected effects. The system is applied to the the task of controlling the growth of plants, where the evolution of the plant attributes strongly depends on different events taking place in the temporally distant past history of the plant. The main problem to tackle is how to efficiently detect these past events. This is very challenging since the inclusion of time could make the dimensionality of the search space extremely large and the collected training instances may only provide very limited information about the relevant combinations of events. To address this problem we propose a learning method that progressively identifies those events that are more likely to produce a sequence of changes under a plant treatment. Since the number of experiences is very limited compared to the size of the event space, we use a probabilistic estimate that takes into account the lack of experience to prevent biased estimations. Planning operators are generated from most accurately predicted sequences of changes. Planning and learning are integrated in a decision-making framework that operates without task interruptions by allowing a human gardener to instruct the treatments when the knowledge acquired so far is not enough to make a decision.

Abstract Understanding and learning the semantics of complex manipulation actions are intriguing and non-trivial issues for the development of autonomous robots. In this paper, we present a novel method for an on-line, incremental learning of the semantics of manipulation actions by observation. Recently, we had introduced the Semantic Event Chains (SECs) as a new generic representation for manipulations, which can be directly computed from a stream of images and is based on the changes in the relationships between objects involved in a manipulation. We here show that the SEC concept can be used to bootstrap the learning of the semantics of manipulation actions without using any prior knowledge about actions or objects. We create a new manipulation action benchmark with 8 different manipulation tasks including in total 120 samples to learn an archetypal SEC model for each manipulation action. We then evaluate the learned SEC models with 20 long and complex chained manipulation sequences including in total 103 manipulation samples. Thereby we put the event chains to a decisive test asking how powerful is action classification when using this framework. We find that we reach up to 100 % and 87 % average precision and recall values in the validation phase and 99 % and 92 % in the testing phase. This supports the notion that SECs are a useful tool for classifying manipulation actions in a fully automatic way.

An originally chaotic system can be controlled into various periodic dynamics. When it is implemented into a legged robots locomotion control as a central pattern generator CPG, sophisticated gait patterns arise so that the robot can perform various walking behaviors. However, such a single chaotic CPG controller has difficulties dealing with leg malfunction. Specifically, in the scenarios presented here, its movement permanently deviates from the desired trajectory. To address this problem, we extend the single chaotic CPG to multiple CPGs with learning. The learning mechanism is based on a simulated annealing algorithm. In a normal situation, the CPGs synchronize and their dynamics are identical. With leg malfunction or disability, the CPGs lose synchronization leading to independent dynamics. In this case, the learning mechanism is applied to automatically adjust the remaining legs oscillation frequencies so that the robot adapts its locomotion to deal with the malfunction. As a consequence, the trajectory produced by the multiple chaotic CPGs resembles the original trajectory far better than the one produced by only a single CPG. The performance of the system is evaluated first in a physical simulation of a quadruped as well as a hexapod robot and finally in a real six-legged walking machine called AMOSII. The experimental results presented here reveal that using multiple CPGs with learning is an effective approach for adaptive locomotion generation where, for instance, different body parts have to perform independent movements for malfunction compensation

The neuromechanical control principles of animal locomotion provide good insights for the development of bio-inspired legged robots for walking on challenging surfaces. Based on such principles, we developed a neuromechanical controller consisting of a modular neural network (MNN) and of virtual agonist-antagonist muscle mechanisms (VAAMs). The controller allows for variable compliant leg motions of a hexapod robot, thereby leading to energy-efficient walking on different surfaces. Without any passive mechanisms or torque and position feedback at each joint, the variable compliant leg motions are achieved by only changing the stiffness parameters of the VAAMs. In addition, six surfaces can be also classified by observing the motor signals generated by the controller. The performance of the controller is tested on a physical hexapod robot. Experimental results show that it can effectively walk on six different surfaces with the specific resistances between 9.1 and 25.0, and also classify them with high accuracy.

The problem of how to arrive at an appropriate 3D-segmentation of a scene remains difficult. While current state-of-the-art methods continue to gradually improve in benchmark performance, they also grow more and more complex, for example by incorporating chains of classifiers, which require training on large manually annotated data- sets. As an alternative to this, we present a new, efficient learning- and model-free approach for the segmentation of 3D point clouds into object parts. The algorithm begins by decomposing the scene into an adjacency-graph of surface patches based on a voxel grid. Edges in the graph are then classified as either convex or concave using a novel combination of simple criteria which operate on the local geometry of these patches. This way the graph is divided into locally convex connected subgraphs, which - with high accuracy - represent object parts. Additionally, we propose a novel depth dependent voxel grid to deal with the decreasing point-density at far distances in the point clouds. This improves segmentation, allowing the use of fixed parameters for vastly different scenes. The algorithm is straight-forward to implement and requires no training data, while nevertheless producing results that are comparable to state-of-the-art methods which incorporate high-level concepts involving classification, learning and model fitting.

In this paper we apply a policy improvement algorithm called Policy Improvement with Path Integrals (PI2) to generate goal-directed locomotion of a complex snake-like robot with screw-drive units. PI2 is numerically simple and has an ability to deal with high dimensional systems. Here, this approach is used to find proper locomotion control parameters, like joint angles and screw-drive velocities, of the robot. The learning process was achieved using a simulated robot and the learned parameters were successfully transferred to the real one. As a result the robot can locomote toward a given goal.

Walking animals show fascinating locomotor abilities and complex behaviors. Biological study has revealed that such complex behaviors is a result of a combination of biomechanics and neural mechanisms. While biomechanics allows for flexibility and a variety of movements, neural mechanisms generate locomotion, make predictions, and provide adaptation. Inspired by this finding, we present here an artificial bio-inspired walking system which combines biomechanics (in terms of its body and leg structures) and neural mechanisms. The neural mechanisms consist of 1) central pattern generator-based control for generating basic rhythmic patterns and coordinated movements, 2) reservoir-based adaptive forward models with efference copies for sensory prediction as well as state estimation, and 3) searching and elevation control for adapting the movement of an individual leg to deal with different environmental conditions. Simulation results show that this bio-inspired approach allows the walking robot to perform complex locomotor abilities including walking on undulated terrains, crossing a large gap, as well as climbing over a high obstacle and a fleet of stairs.

In this article, we describe technologies facilitating the set-up of automated assembly solutions which have been developed in the context of the IntellAct project (2011-2014). Tedious procedures are currently still required to establish such robot solutions. This hinders especially the automation of so called few-of-a-kind production. Therefore, most production of this kind is done manually and thus often performed in low-wage countries. In the IntellAct project, we have developed a set of methods which facilitate the set-up of a complex automatic assembly process, and here we present our work on tele-operation, dexterous grasping, pose estimation and learning of control strategies. The prototype developed in IntellAct is at a TRL4 (corresponding to demonstration in lab environment).

The proposed framework addresses the problem of implementing a high level

Increasingly widespread available haptic sensors mounted on articulated hands offer new sensory channels that can complement shape extraction from vision to enable a more robust handling of objects in cases when vision is restricted or even unavailable. However, to estimate object shape from haptic interaction data is a difficult challenge due to the complexity of the contact interaction between the movable object and sensor surfaces, leading to a coupled estimation problem of shape and object pose. While for vision efficient solutions to the underlying SLAM problem are known, the available information is much sparser in the tactile case, posing great difficulties for a straightforward adoption of standard SLAM algorithms. In the present paper, we thus explore whether a biologically inspired model based on dynamic neural fields can offer a route towards a practical algorithm for tactile SLAM. Our study is focused on a restricted scenario where a two-fingered robot hand manipulates an n-gon with a fixed rotational axis. We demonstrate that our model can accumulate shape information from reasonably short interaction sequences and autonomously build a representation despite significant ambiguity of the tactile data due to the rotational periodicity of the object. We conclude that the presented framework may be a suitable basis to solve the tactile SLAM problem also in more general settings which will be the focus of subsequent work.

Purpose - Biological muscles of animals have a surprising variety of functions, i.e., struts, springs, and brakes. According to this, the purpose of this paper is to apply virtual agonist-antagonist mechanisms to robot joint control allowing for muscle-like functions and variably compliant joint motions. Design/methodology/approach - Each joint is driven by a pair of virtual agonist-antagonist mechanism (VAAM, i.e., passive components). The muscle-like functions as well as the variable joint compliance are simply achieved by tuning the damping coefficient of the VAAM. Findings - With the VAAM, variably compliant joint motions can be produced without mechanically bulky and complex mechanisms or complex force/toque sensing at each joint. Moreover, through tuning the damping coefficient of the VAAM, the functions of the VAAM are comparable to biological muscles. Originality/value - The model (i.e., VAAM) provides a way forward to emulate muscle-like functions that are comparable to those found in physiological experiments of biological muscles. Based on these muscle-like functions, the robotic joints can easily achieve variable compliance that does not require complex physical components or torque sensing systems thereby capable of implementing the model on small legged robots driven by, e.g., standard servo motors. Thus, the VAAM minimizes hardware and reduces system complexity. From this point of view, the model opens up another way of simulating muscle behaviors on artificial machines. Executive summary The VAAM can be applied to produce variable compliant motions of a high DOF robot. Only relying on force sensing at the end effector, this application is easily achieved by changing coefficients of the VAAM. Therefore, the VAAM can reduce economic cost on mechanical and sensing components of the robot, compared to traditional methods (e.g., artificial muscles).

Self-organization, especially in the framework of embodiment in biologically inspired robots, allows the acquisition of behavioral primitives by autonomous robots themselves. However, it is an open question how self-organization of basic motor primitives and goal-orientation can be combined, which is a prerequisite for the usefulness of such systems. In the paper at hand we propose a goal-orientation framework allowing the combination of self-organization and goal-orientation for the control of autonomous robots in a mutually independent fashion. Self-organization based motor primitives are employed to achieve a given goal. This requires less initial knowledge about the properties of robot and environment and increases adaptivity of the overall system. A combination of self-organization and reward-based learning seems thus a promising route for the development of adaptive learning systems.

Physiological studies suggest that the integration of neural circuits and biomechanics (e.g., muscles) is a key for animals to achieve robust and efficient locomotion over challenging surfaces. Inspired by these studies, we present a neuromechanical controller of a hexapod robot for walking on soft elastic and loose surfaces. It consists of a modular neural network (MNN) and virtual agonist-antagonist mechanisms (VAAM, i.e., a muscle model). The MNN coordinates 18 joints and generates basic locomotion while variable joint compliance for walking on different surfaces is achieved by the VAAM. The changeable compliance of each joint does not depend on physical compliant mechanisms or joint torque sensing. Instead, the compliance is altered by two internal parameters of the VAAM. The performance of the controller is tested on a physical hexapod robot for walking on soft elastic (e.g., sponge) and loose (e.g., gravel and snow) surfaces. The experimental results show that the controller enables the hexapod robot to achieve variably compliant leg behaviors, thereby leading to more energy-efficient locomotion on different surfaces. In addition, a finding of the experiments complies with the finding of physiological experiments on cockroach locomotion on soft elastic surfaces.

The segmentation of textured and noisy areas in images is a very challenging task due to the large variety of objects and materials in natural environments, which cannot be solved by a single similarity measure. In this paper, we address this problem by proposing a novel edge-preserving texture filter, which smudges the color values inside uniformly textured areas, thus making the processed image more workable for color-based image segmentation. Due to the highly parallel structure of the method, the implementation on a GPU runs in real-time, allowing us to process standard images within tens of milliseconds. By preprocessing images with this novel filter before applying a recent real-time color-based image segmentation method, we obtain significant improvements in performance for images from the Berkeley dataset, outperforming an alternative version using a standard bilateral filter for preprocessing. We further show that our combined approach leads to better segmentations in terms of a standard performance measure than graph-based and mean-shift segmentation for the Berkeley image dataset.

Humans can perform a multitude of different actions with their hands (manipulations). In spite of this, so far there have been only a few attempts to represent manipulation types trying to understand the underlying principles. Here we first discuss how manipulation actions are structured in space and time. For this we use as temporal anchor points those moments where two objects (or hand and object) touch or un-touch each other during a manipulation. We show that by this one can define a relatively small tree-like manipulation ontology. We find less than 30 fundamental manipulations. The temporal anchors also provide us with information about when to pay attention to additional important information, for example when to consider trajectory shapes and relative poses between objects. As a consequence a highly condensed representation emerges by which different manipulations can be recognized and encoded. Examples of manipulations recognition and execution by a robot based on this representation are given at the end of this study.

Physiological studies suggest that the integration of neural circuits and biomechanics (e.g., muscles) is a key for animals to achieve robust and efficient locomotion over challenging surfaces. Inspired by these studies, we present a neuromechanical controller of a hexapod robot for walking on soft elastic and loose surfaces. It consists of a modular neural network (MNN) and virtual agonist-antagonist mechanisms (VAAM, i.e., a muscle model). The MNN coordinates 18 joints and generates basic locomotion while variable joint compliance for walking on different surfaces is achieved by the VAAM. The changeable compliance of each joint does not depend on physical compliant mechanisms or joint torque sensing. Instead, the compliance is altered by two internal parameters of the VAAM. The performance of the controller is tested on a physical hexapod robot for walking on soft elastic (e.g., sponge) and loose (e.g., gravel and snow) surfaces. The experimental results show that the controller enables the hexapod robot to achieve variably compliant leg behaviors, thereby leading to more energy-efficient locomotion on different surfaces. In addition, a finding of the experiments complies with the finding of physiological experiments on cockroach locomotion on soft elastic surfaces,

Understanding the exact mechanism of learning and memory emerging from complex dynamical systems like neural networks serves as a challenging field of research. Traditionally the neural mechanisms underlying memory and cognition in these systems are described by steady-state or stable fixed point attractor dynamics. However an alternative and refined understanding of the neuronal dynamics can be achieved through the idea of transient dynamics 1 (reservoir computing paradigm) i.e., computation through input specific trajectories in neural space without stable equilibrium. Mathematical analysis of the underlying memory through such transient dynamics is difficult. As such information theory provides tools to quantify the dynamics of memory in such networks. One such popular measure of memory capacity in reservoir networks is the linear memory capacity 2. It provides an indication of how well the network can reconstruct delayed versions of the input signal. However it assumes a linear retrieval of input signal and deteriorates with neuron non-linearity. Alternatively, active information storage 3 provides a measure of local neuron memory by quantifying the degree of influence of past activity on the next time step activity of a neuron independent of neuronal non-linearity. In this work we further extend this quantity by calculating the mutual information between a neuron past activity and its immediate future activity while conditioning out delayed versions of the input signal. Summing over different delays of input signal it provides a suitable measure of total input driven active memory in the network. Intuitively active memory calculates the actual memory in use i.e. influence of input history on local neuron memory. We compare memory capacity and active memory (AM) with different network parameters for networks driven with statistically different inputs and justify AM as an appropriate means to quantify the dynamics of memory in input driven neural networks.

Central pattern generators (CPGs) play a crucial role for animal locomotion control. They can be entrained by sensory feedback to induce proper rhythmic patterns and even store the entrained patterns through connection weights. Inspired by this biological finding, we use four adaptive neural oscillators with synaptic plasticity as CPGs for locomotion control of our real snake-like robot with screw-drive mechanism. Each oscillator consists of only three neurons and uses adaptive mechanisms based on frequency adaptation and Hebbian-type learning rules. It autonomously generates proper periodic patterns for the robot locomotion and can be entrained by sensory feedback to memorize the patterns. The adaptive CPG system in conjunction with a simple control strategy enables the robot to perform self-tuning behavior which is robust against short-time perturbations. The generated behavior is also energy efficient. In addition, the robot can also cope with corners as well as move through a complex environment with obstacles.

Fruit flies Drosophila melanogaster rely on their olfactory system to process environmental information. This information has to be transmitted without system-relevant loss by the olfactory system to deeper brain areas for learning. Here we study the role of several parameters of the flys olfactory system and the environment and how they influence olfactory information transmission. We have designed an abstract model of the antennal lobe, the mushroom body and the inhibitory circuitry. Mutual information between the olfactory environment, simulated in terms of different odor concentrations, and a sub-population of intrinsic mushroom body neurons Kenyon cells was calculated to quantify the efficiency of information transmission. With this method we study, on the one hand, the effect of different connectivity rates between olfactory projection neurons and firing thresholds of Kenyon cells. On the other hand, we analyze the influence of inhibition on mutual information between environment and mushroom body. Our simulations show an expected linear relation between the connectivity rate between the antennal lobe and the mushroom body and firing threshold of the Kenyon cells to obtain maximum mutual information for both low and high odor concentrations. However, contradicting all-day experiences, high odor concentrations cause a drastic, and unrealistic, decrease in mutual information for all connectivity rates compared to low concentration. But when inhibition on the mushroom body is included, mutual information remains at high levels independent of other system parameters. This finding points to a pivotal role of inhibition in fly information processing without which the systems efficiency will be substantially reduced

Classical conditioning (conventionally modeled as correlation-based learning) and operant conditioning (conventionally modeled as reinforcement learning or reward-based learning) have been found in biological systems. Evidence shows that these two mechanisms strongly involve learning about associations. Based on these biological findings, we propose a new learning model to achieve successful control policies for artificial systems. This model combines correlation-based learning using input correlation learning (ICO learning) and reward-based learning using continuous actor-critic reinforcement learning (RL), thereby working as a dual learner system. The model performance is evaluated by simulations of a cart-pole system as a dynamic motion control problem and a mobile robot system as a goal-directed behavior control problem. Results show that the model can strongly improve pole balancing control policy, i.e., it allows the controller to learn stabilizing the pole in the largest domain of initial conditions compared to the results obtained when using a single learning mechanism. This model can also find a successful control policy for goal-directed behavior, i.e., the robot can effectively learn to approach a given goal compared to its individual components. Thus, the study pursued here sharpens our understanding of how two different learning mechanisms can be combined and complement each other for solving complex tasks.

Recurrent neural networks of the reservoir computing (RC) type have been found useful in various time-series processing tasks with inherent non-linearity and requirements of variable temporal memory. Specifically for delayed response tasks involving the transient memorization of information (temporal memory), self-adaptation in RC is crucial for generalization to varying delays. In this work using information theory, we combine a generalized intrinsic plasticity rule with a local information dynamics based schema of reservoir neuron leak adaptation. This allows the RC network to be optimized in a self-adaptive manner with minimal parameter tuning. Local active information storage, measured as the degree of influence of previous activity on the next time step activity of a neuron, is used to modify its leak-rate. This results in RC network with non-uniform leak rate which depends on the time scales of the incoming input. Intrinsic plasticity (IP) is aimed at maximizing the mutual information between each neurons input and output while maintaining a mean level of activity (homeostasis). Experimental results on two standard benchmark tasks confirm the extended performance of this system as compared to the static RC (fixed leak and no IP) and RC with only IP. In addition, using both a simulated wheeled robot and a more complex physical hexapod robot, we demonstrate the ability of the system to achieve long temporal memory for solving a basic T-shaped maze navigation task with varying delay time scale.

Since several years dynamic movement primitives (DMPs) are more and more getting into the center of interest for flexible movement control in robotics. In this study we introduce sensory feedback together with a predictive learning mechanism which allows tightly coupled dual-agent systems to learn an adaptive, sensor-driven interaction based on DMPs. The coupled conventional (no-sensors, no learning) DMP-system automatically equilibrates and can still be solved analytically allowing us to derive conditions for stability. When adding adaptive sensor control we can show that both agents learn to cooperate. Simulations as well as real-robot experiments are shown. Interestingly, all these mechanisms are entirely based on low level interactions without any planning or cognitive component.

Learning of goal-directed behaviors in biological systems is broadly based on associations between conditional and unconditional stimuli. This can be further classified as classical conditioning (correlation-based learning) and operant conditioning (reward-based learning). Although traditionally modeled as separate learning systems in artificial agents, numerous animal experiments point towards their co-operative role in behavioral learning. Based on this concept, the recently introduced framework of neural combinatorial learning combines the two systems where both the systems run in parallel to guide the overall learned behavior. Such a combinatorial learning demonstrates a faster and efficient learner. In this work, we further improve the framework by applying a reservoir computing network (RC) as an adaptive critic unit and reward modulated Hebbian plasticity. Using a mobile robot system for goal-directed behavior learning, we clearly demonstrate that the reservoir critic outperforms traditional radial basis function (RBF) critics in terms of stability of convergence and learning time. Furthermore the temporal memory in RC allows the system to learn partially observable markov decision process scenario, in contrast to a memory less RBF critic.

Living creatures, like walking animals, have found fascinating solutions for the problem of locomotion control. Their movements show the impression of elegance including versatile, energy-efficient, and adaptable locomotion. During the last few decades, roboticists have tried to imitate such natural properties with artificial legged locomotion systems by using different approaches including machine learning algorithms, classical engineering control techniques, and biologically-inspired control mechanisms. However, their levels of performance are still far from the natural ones. By contrast, animal locomotion mechanisms seem to largely depend not only on central mechanisms (central pattern generators, CPGs) and sensory feedback (afferent-based control) but also on internal forward models (efference copies). They are used to a different degree in different animals. Generally, CPGs organize basic rhythmic motions which are shaped by sensory feedback while internal models are used for sensory prediction and state estimations. According to this concept, we present here adaptive neural locomotion control consisting of a CPG mechanism with neuromodulation and local leg control mechanisms based on sensory feedback and adaptive neural forward models with efference copies. This neural closed-loop controller enables a walking machine to perform a multitude of different walking patterns including insect-like leg movements and gaits as well as energy-efficient locomotion. In addition, the forward models allow the machine to autonomously adapt its locomotion to deal with a change of terrain, losing of ground contact during stance phase, stepping on or hitting an obstacle during swing phase, leg damage, and even to promote cockroach-like climbing behavior. Thus, the results presented here show that the employed embodied neural closed-loop system can be a powerful way for developing robust and adaptable machines.

This paper utilizes a 2D laser range finder (LRF) to determine the behavior of a walking robot. The LRF provides information for 1) obstacle/gap detection as well as 2) terrain classification. The obstacle/gap detection is based on an edge detection with increased robustness and accuracy due to customized pre and post processing. Its output is used to drive obstacle/gap avoidance behavior or climbing behavior, depending on the height of obstacles or the depth of gaps. The terrain classification employs terrain roughness to select a proper gait with respect to the current terrain. As a result, the combination of these methods enables the robot to decide if obstacles and gaps can be climbed up/down or have to be avoided while at the same time a terrain specific gait can be chosen.

Robust visual tracking is an essential precursor to understanding and replicating human actions in robotic systems. In order to accurately evaluate the semantic meaning of a sequence of video frames, or to replicate an action contained therein, one must be able to coherently track and segment all observed agents and objects. This work proposes a novel online point cloud based algorithm which simultaneously tracks 6DoF pose and determines spatial extent of all entities in indoor scenarios. This is accomplished using a persistent supervoxel world-model which is updated, rather than replaced, as new frames of data arrive. Maintenance of a world model enables general object permanence, permitting successful tracking through full occlusions. Object models are tracked using a bank of independent adaptive particle filters which use a supervoxel observation model to give rough estimates of object state. These are united using a novel multi-model RANSAC-like approach, which seeks to minimize a global energy function associating world-model supervoxels to predicted states. We present results on a standard robotic assembly benchmark for two application scenarios - human trajectory imitation and semantic action understanding - demonstrating the usefulness of the tracking in intelligent robotic systems.

Generalization in robotics is one of the most important problems. New generalization approaches use internet databases in order to solve new tasks. Modern search engines can return a large amount of information according to a query within milliseconds. However, not all of the returned information is task relevant, partly due to the problem of polysemes. Here we specifically address the problem of object generalization by using image search. We suggest a bi-modal solution, combining visual and textual information, based on the observation that humans use additional linguistic cues to demarcate intended word meaning. We evaluate the quality of our approach by comparing it to human labelled data and find that, on average, our approach leads to improved results in comparison to Google searches, and that it can treat the problem of polysemes.

The paper presents research results, showing how unsupervised and supervised ontology learning methods can be combined in an action ontology building approach. A framework for action ontology building from domain-specific corpus texts is suggested, using different natural language processing techniques, such as collocation extraction, frequency lists, word space model, etc. The suggested framework employs additional knowledge sources of WordNet and VerbNet with structured linguistic and semantic information. Re-sults from experiments with crawled chemical laboratory corpus texts are given

In this paper, we investigated the dynamics of a hexapod robot model whose legs are driven by nonlinear oscillators with a phase modulation mechanism including phase resetting and inhibition. This mechanism changes the oscillation period of the oscillator depending solely on the timing of the foots contact. This strategy is based on observation of animals. The performance of the controller is evaluated using a physical simulation environment. Our simulation results show that the robot produces some stable gaits depending on the locomotion speed due to the phase modulation mechanism, which are simillar to the gaits of insects.

Autonomous robots are faced with the problem of encoding complex actions (e.g. complete manipulations) in a generic and generalizable way. Recently we had introduced the Semantic Event Chains (SECs) as a new representation which can be directly computed from a stream of 3D images and is based on changes in the relationships between objects involved in a manipulation. Here we show that the SEC framework can be extended (called extended SEC) with action-related information and used to achieve and encode two important cognitive properties relevant for advanced autonomous robots: The extended SEC enables us to determine whether an action representation (1) needs to be newly created and stored in its entirety in the robots memory or (2) whether one of the already known and memorized action representations just needs to be refined. In human cognition these two processes (1 and 2) are known as accommodation and assimilation. Thus, here we show that the extended SEC representation can be used to realize these processes originally defined by Piaget for the first time in a robotic application. This is of fundamental importance for any cognitive agent as it allows categorizing observed actions in new versus known ones, storing only the relevant aspects.

Memory storage in the brain relies on mechanisms acting on time scales from minutes, for long-term synaptic potentiation, to days, for memory consolidation. During such processes, neural circuits distinguish synapses relevant for forming a long-term storage, which are consolidated, from synapses of short-term storage, which fade. How time scale integration and synaptic differentiation is simultaneously achieved remains unclear. Here we show that synaptic scaling - a slow process usually associated with the maintenance of activity homeostasis - combined with synaptic plasticity may simultaneously achieve both, thereby providing a natural separation of short- from long-term storage. The interaction between plasticity and scaling provides also an explanation for an established paradox where memory consolidation critically depends on the exact order of learning and recall. These results indicate that scaling may be fundamental for stabilizing memories, providing a dynamic link between early and late memory formation processes.

The goal of this study is to provide an architecture for a generic definition of robot manipulation actions. We emphasize that the representation of actions presented here is procedural. Thus, we will define the structural elements of our action representations as execution protocols. To achieve this, manipulations are defined using three levels. The top- level defines objects, their relations and the actions in an abstract and symbolic way. A mid-level sequencer, with which the action primitives are chained, is used to structure the actual action execution, which is performed via the bottom level. This (lowest) level collects data from sensors and communicates with the control system of the robot. This method enables robot manipulators to execute the same action in different situations i.e. on different objects with different positions and orientations. In addition, two methods of detecting action failure are provided which are necessary to handle faults in system. To demonstrate the effectiveness of the proposed framework, several different actions are performed on our robotic setup and results are shown. This way we are creating a library of human-like robot actions, which can be used by higher-level task planners to execute more complex tasks.

Animals can move in not only elegant but also energy efficient ways. Their skin is one of the key components for this achievement. It provides a proper friction for forward motion and can protect them from slipping on a surface during locomotion. Inspired by this, we applied real shark skin to the foot soles of our hexapod robot AMOS. The material is formed to cover each foot of AMOS. Due to shark skin texture which has asymmetric profile inducing frictional anisotropy, this feature allows AMOS to grip specific surfaces and effectively locomote without slipping. Using real-time walking experiments, this study shows that implementing the biological material on the robot can reduce energy consumption while walking up a steep slope covered by carpets or other felt-like or rough substrates.

Legged robots need to be able to classify and recognize different terrains to adapt their gait accordingly. Recent works in terrain classification use different types of sensors (like stereovision, 3D laser range, and tactile sensors) and their combination. However, such sensor systems require more computing power, produce extra load to legged robots, and/or might be difficult to install on a small size legged robot. In this work, we present an online terrain classification system. It uses only a monocular camera with a feature-based terrain classification algorithm which is robust to changes in illumination and view points. For this algorithm, we extract local features of terrains using either Scale Invariant Feature Transform (SIFT) or Speed Up Robust Feature (SURF). We encode the features using the Bag of Words (BoW) technique, and then classify the words using Support Vector Machines (SVMs) with a radial basis function kernel. We compare this feature-based approach with a color-based approach on the Caltech-256 benchmark as well as eight different terrain image sets (grass, gravel, pavement, sand, asphalt, floor, mud, and fine gravel). For terrain images, we observe up to 90% accuracy with the feature-based approach. Finally, this online terrain classification system is successfully applied to our small hexapod robot AMOS II. The output of the system providing terrain information is used as an input to its neural locomotion control to trigger an energy-efficient gait while traversing different terrains.

Unsupervised over-segmentation of an image into regions of perceptually similar pixels, known as super pixels, is a widely used preprocessing step in segmentation algorithms. Super pixel methods reduce the number of regions that must be considered later by more computationally expensive algorithms, with a minimal loss of information. Nevertheless, as some information is inevitably lost, it is vital that super pixels not cross object boundaries, as such errors will propagate through later steps. Existing methods make use of projected color or depth information, but do not consider three dimensional geometric relationships between observed data points which can be used to prevent super pixels from crossing regions of empty space. We propose a novel over-segmentation algorithm which uses voxel relationships to produce over-segmentations which are fully consistent with the spatial geometry of the scene in three dimensional, rather than projective, space. Enforcing the constraint that segmented regions must have spatial connectivity prevents label flow across semantic object boundaries which might otherwise be violated. Additionally, as the algorithm works directly in 3D space, observations from several calibrated RGB+D cameras can be segmented jointly. Experiments on a large data set of human annotated RGB+D images demonstrate a significant reduction in occurrence of clusters crossing object boundaries, while maintaining speeds comparable to state-of-the-art 2D methods.

We present an architecture for real-time, online vision systems which enables development and use of complex vision pipelines integrating any number of algorithms. Individual algorithms are implemented using modular plugins, allowing integration of independently developed algorithms and rapid testing of new vision pipeline configurations. The architecture exploits the parallelization of graphics processing units (GPUs) and multi-core systems to speed processing and achieve real-time performance. Additionally, the use of a global memory management system for frame buffering permits complex algorithmic flow (e.g. feedback loops) in online processing setups, while maintaining the benefits of threaded asynchronous operation of separate algorithms. To demonstrate the system, a typical real-time system setup is described which incorporates plugins for video and depth acquisition, GPU-based segmentation and optical flow, semantic graph generation, and online visualization of output. Performance numbers are shown which demonstrate the insignificant overhead cost of the architecture as well as speed-up over strictly CPU and single threaded implementations.

Humans can perform a multitude of different actions with their hands (manipulations). In spite of this, so far there have been only a few attempts to represent manipulation types trying to understand the underlying principles. Here we first discuss how manipulation actions are structured in space and time. For this we use as temporal anchor points those moments where two objects (or hand and object) touch or un-touch each other during a manipulation. We show that by this one can define a relatively small tree-like manipulation ontology. We find less than 30 fundamental manipulations. The temporal anchors also provide us with information about when to pay attention to additional important information, for example when to consider trajectory shapes and relative poses between objects. As a consequence a highly condensed representation emerges by which different manipulations can be recognized and encoded. Examples of manipulations recognition and execution by a robot based on this representation are given at the end of this study.

Rhythmic neural circuits play an important role in biological systems in particular in motion generation. They can be entrained by sensory feedback to induce rhythmic motion at a natural frequency, leading to energy-efficient motion. In addition, such circuits can even store the entrained rhythmical patterns through connection weights. Inspired by this, we introduce an adaptive discrete-time neural oscillator system with synaptic plasticity. The system consists of only three neurons and uses adaptive mechanisms based on frequency adaptation and Hebbian-type learning rules. As a result, it autonomously generates periodic patterns and can be entrained by sensory feedback to memorize a pattern. Using numerical simulations we show that this neural system possesses fast and precise convergence behaviour within a wide target frequency range. We use resonant tuning of a pendulum as a simple system for demonstrating possible applications of the adaptive oscillator network.

Recently, an integrative view of neural circuits and mechanical components has been developed by neuroscientists and biomechanicians 11, 8. This view argues that mechanical components cannot be isolated from neural circuits in the context of substantially perturbed locomotion. Note that mechanical passive walkers with no neural circuits only show stable locomotion on flat terrain or small slopes 2. The argument of the integrative view has been supported by a cockroach experiment, which has demonstrated that more modulations of neural activities are detected when cockroaches run over a highly complex terrain with larger obstacles (more than three times cockroach hip height). Normally, cockroaches are able to solely rely on passive mechanical properties for rapid stabilization while confronted with moderate obstacles (less than three times cockroach hip height) 10. In addition, neural circuits and leg muscle activities tend to be entrained by mechanical feedback 11, 12, 14. Besides, it is well known that neural activities modulate muscle impedance such as stiffness and damping 7, 9, 15, such modulations can be utilized for stabilization in posture and locomotion 3.

Conventional synaptic plasticity in combination with synaptic scaling is a biologically plau-sible plasticity rule that guides the development of synapses toward stability. Here we analyze the development of synaptic connections and the resulting activity patterns in dif-ferent feed-forward and recurrent neural networks, with plasticity and scaling. We show under which constraints an external input given to a feed-forward network forms an input trace similar to a cell assembly Hebb, 1949 by enhancing synaptic weights to larger stable values as compared to the rest of the network. For instance, a weak input creates a less strong representation in the network than a strong input which produces a trace along large parts of the network.These processes are strongly influenced by the underlying con-nectivity. For example, when embedding recurrent structures excitatory rings, etc. into a feed-forward network, the input trace is extended into more distant layers, while inhibition shortens it. These findings provide a better understanding of the dynamics of generic net-work structures where plasticity is combined with scaling. This makes it also possible to use this rule for constructing an artificial network with certain desired storage properties

Insects, e.g. cockroaches and stick insects, have found fascinating solutions for the problem of locomotion, especially climbing over a large variety of obstacles. Research on behavioral neurobiology has identified key behavioral patterns of these animals (i.e., body flexion, center of mass elevation, and local leg reflexes) necessary for climbing. Inspired by this finding, we develop a neural control mechanism for hexapod robots which generates basic walking behavior and especially enables them to effectively perform reactive climbing behavior. The mechanism is composed of three main neural circuits: locomotion control, reactive backbone joint control, and local leg reflex control. It was developed and tested using a physical simulation environment, and was then successfully transferred to a physical six-legged walking machine, called AMOS II. Experimental results show that the controller allows the robot to overcome obstacles of various heights (e.g., 75% of its leg length, which are higher than those that other comparable legged robots have achieved so far). The generated climbing behavior is also comparable to the one observed in cockroaches.

We present a real-time technique for the spatiotemporal segmentation of color/depth movies. Images are segmented using a parallel Metropolis algorithm implemented on a GPU utilizing both color and depth information, acquired with the Microsoft Kinect. Segments represent the equilibrium states of a Potts model, where tracking of segments is achieved by warping obtained segment labels to the next frame using real-time optical flow, which reduces the number of iterations required for the Metropolis method to encounter the new equilibrium state. By including depth information into the framework, true objects boundaries can be found more easily, improving also the temporal coherency of the method. The algorithm has been tested for videos of medium resolutions showing human manipulations of objects. The framework provides an inexpensive visual front end for visual preprocessing of videos in industrial settings and robot labs which can potentially be used in various applications.

The firing of hippocampal place cells encodes instantaneous location but can also reflect where the animal is heading prospective firing, or where it has just come from retrospective firing. The current experiment sought to explicitly control the prospective firing of place cells with a visual discriminada in a T-maze. Rats were trained to associate a specific visual stimulus e.g. a flashing light with the occurrence of reward in a specific location e.g. and the left arm of the T. A different visual stimulus e.g. and a constant light signalled the availability of reward in the opposite arm of the T. After this discrimination had been acquired, rats were implanted with electrodes in the CA1 layer of the hippocampus. Place cells were then identified and recorded as the animals performed the discrimination task, and the presentation of the visual stimulus was manipulated. A subset of CA1 place cells fired at different rates on the central stem of the T depending on the animals intended destination, but this conditional or prospective firing was independent of the visual discriminative stimulus. The firing rate of some place cells was, however, modulated by changes in the timing of presentation of the visual stimulus. Thus, place cells fired prospectively, but this firing did not appear to be controlled, directly, by a salient visual stimulus that controlled behaviour

Recurrent neural networks of the Reservoir Computing (RC) type have been found useful in various time-series processing tasks with inherent non-linearity and requirements of temporal memory. Here with the aim to obtain extended temporal memory in generic delayed response tasks, we combine a generalised intrinsic plasticity mechanism with an information storage based neuron leak adaptation rule in a self-organised manner. This results in adaptation of neuron local memory in terms of leakage along with inherent homeostatic stability. Experimental results on two benchmark tasks confirm the extended performance of this system as compared to a static RC and RC with only intrinsic plasticity. Furthermore, we demonstrate the ability of the system to solve long temporal memory tasks via a simulated T-shaped maze navigation scenario.

The generation of complex movement patterns, in particular in cases where one needs to smoothly and accurately join trajectories in a dynamic way, is an important problem in robotics. This paper presents a novel joining method based on the modification of the original dynamic movement primitive DMP formulation. The new method can reproduce the target trajectory with high accuracy regarding both, position and velocity profile, and produces smooth and natural transitions in position as well as velocity space. The properties of the method are demonstrated by applying it to simulated handwriting generation also shown on a robot, where an adaptive algorithm is used to learn trajectories from human demonstration. These results demonstrate that the new method is a feasible alternative for joining of movement sequences which has high potential for all robotics applications where trajectory joining is required

In chaos control, an originally chaotic system is modified so that periodic dynamics arise. One application of this is to use the periodic dynamics of a single chaotic system as walking patterns in legged robots. In our previous work we applied such a controlled chaotic system as a central pattern generator (CPG) to generate different gait patterns of our hexapod robot AMOSII. However, if one or more legs break, its control fails. Specifically, in the scenario presented here, its movement permanently deviates from a desired trajectory. This is in contrast to the movement of real insects as they can compensate for body damages, for instance, by adjusting the remaining legs frequency. To achieve this for our hexapod robot, we extend the system from one chaotic system serving as a single CPG to multiple chaotic systems, performing as multiple CPGs. Without damage, the chaotic systems synchronize and their dynamics is identical (similar to a single CPG). With damage, they can lose synchronization leading to independent dynamics. In both simulations and real experiments, we can tune the oscillation frequency of every CPG manually so that the controller can indeed compensate for leg damage. In comparison to the trajectory of the robot controlled by only a single CPG, the trajectory produced by multiple chaotic CPG controllers resembles the original trajectory by far better. Thus, multiple chaotic systems that synchronize for normal behavior but can stay desynchronized in other circumstances are an effective way to control complex behaviors where, for instance, different body parts have to do independent movements like after leg damage.

We present a method for unsupervised on-line dense video segmentation which utilizes sequential Bayesian estimation techniques to resolve partial and full occlusions. Consistent labeling through occlusions is vital for applications which move from low-level object labels to high-level semantic knowledge - tasks such as activity recognition or robot control. The proposed method forms a predictive loop between segmentation and tracking, with tracking predictions used to seed the segmentation kernel, and segmentation results used to update tracked models. All segmented labels are tracked, without the use of a-priori models, using parallel color-histogram particle filters. Predictions are combined into a probabilistic representation of image labels, a realization of which is used to seed segmentation. A simulated annealing relaxation process allows the realization to converge to a minimal energy segmented image. Found segments are subsequently used to repopulate the particle sets, closing the loop. Results on the Cranfield benchmark sequence demonstrate that the prediction mechanism allows on-line segmentation to maintain temporally consistent labels through partial & full occlusions, significant appearance changes, and rapid erratic movements. Additionally, we show that tracking performance matches state-of-the art tracking methods on several challenging benchmark sequences.

In mobile robotic applications, visual information needs to be processed fast despite resource limitations of the mobile system. Here, a novel real-time framework for model-free spatiotemporal segmentation of stereo videos is presented. It combines real-time optical flow and stereo with image segmentation and runs on a portable system with an integrated mobile graphics processing unit. The system performs online, automatic, and dense segmentation of stereo videos and serves as a visual front end for preprocessing in mobile robots, providing a condensed representation of the scene that can potentially be utilized in various applications, e.g., object manipulation, manipulation recognition, visual servoing. The method was tested on real-world sequences with arbitrary motions, including videos acquired with a moving camera.

The sigma-point information filters employ a number of deterministic sigma-points to calculate the mean and covariance of a random variable which undergoes a nonlinear transformation. These sigma-points can be generated by the unscented transform or Stirlings interpolation, which corresponds to the unscented information filter (UIF) and the central difference information filter (CDIF) respectively. In this technical note, we develop the square-root extensions of UIF and CDIF, which have better numerical properties than the original versions, e.g., improved numerical accuracy, double order precision and preservation of symmetry. We also show that the square-root unscented information filter (SRUIF) might lose the positive-definiteness due to the negative Cholesky update, whereas the square-root central difference information filter (SRCDIF) has only posi- tive Cholesky update. Therefore, the SRCDIF is preferable to the SRUIF concerning the numerical stability.

In this paper, we present a novel measurement model for particle-filter-based lane shape estimation. Recently, the particle filter has been widely used to solve lane detection and tracking problems, due to its simplicity, robustness, and efficiency. The key part of the particle filter is the measurement model, which describes how well a generated hypothesis (a particle) fits current visual cues in the image. Previous methods often simply combine multiple visual cues in a likelihood function without considering the uncertainties of local visual cues and the accurate probability relationship between visual cues and the lane model. In contrast, this paper derives a new measurement model by utilizing multiple kernel density to precisely estimate this probability relationship. The uncertainties of local visual cues are considered and modeled by Gaussian kernels. Specifically, we use a linear-parabolic model to describe the shape of lane boundaries on a top-view image and a partitioned particle filter (PPF), integrating it with our novel measurement model to estimate lane shapes in consecutive frames. Finally, the robustness of the proposed algorithm with the new measurement model is demonstrated on the DRIVSCO data sets.

This paper presents a new recursive Bayesian estimation method, which is the square-root unscented information filter (SRUIF). The unscented information filter (UIF) has been introduced recently for nonlinear system estimation and sensor fusion. In the UIF framework, a number of sigma points are sampled from the probability distribution of the prior state by the unscented transform and then propagated through the nonlinear dynamic function and measurement function. The new state is estimated from the propagated sigma points. In this way, the UIF can achieve higher estimation accuracies and faster convergence rates than the extended information filter (EIF). As the extension of the original UIF, we propose to use the square-root of the covariance in the SRUIF instead of the full covariance in the UIF for estimation. The new SRUIF has better numerical properties than the original UIF, e.g., improved numerical accuracy, double order precision and preservation of symmetry.

If we stored every bit of input, the storage capacity of our nervous system would be reached after only about 10 days. The nervous system relies on at least two mechanisms that counteract this capacity limit: compression and forgetting. But the latter mechanism needs to know how long an entity should be stored: some memories are relevant only for the next few minutes, some are important even after the passage of several years. Psychology and physiology have found and described many different memory mechanisms, and these mechanisms indeed use different time scales. In this prospect we review these mechanisms with respect to their time scale and propose relations between mechanisms in learning and memory and their underlying physiological basis

In this work we propose an approach for generalization in continuous domain Reinforcement Learning that, instead of using a single function approximator, tries many different function approximators in parallel, each one defined in a different region of the domain. Associated with each approximator is a relevance function that locally quantifies the quality of its approximation, so that, at each input point, the approximator with highest relevance can be selected. The relevance function is defined using parametric estimations of the variance of the q-values and the density of samples in the input space, which are used to quantify the accuracy and the confidence in the approximation, respectively. These parametric estimations are obtained from a probability density distribution represented as a Gaussian Mixture Model embedded in the input-output space of each approximator. In our experiments, the proposed approach required a lesser number of experiences for learning and produced more stable convergence profiles than when using a single function approximator.

In the work at hand we introduce a neural pre- and post-processing framework whose parameters can be adapted by any learning mechanism, e.g. reinforcement learning. The framework allows to generate goal-directed behaviors while at the same time exploiting the bene cial properties, e.g. robustness, of self-organization based primitive behaviors

This paper presents a novel trajectory generator based on Dynamic Movement Primitives DMP. The key ideas from the original DMP formalism are extracted, reformulated and extended from a control theoretical viewpoint. This method can generate smooth trajectories, satisfy position- and velocity boundary conditions at start- and endpoint with high precision, and follow accurately geometrical paths as desired. Paths can be complex and processed as a whole, and smooth transitions can be generated automatically. This novel trajectory generating technology appears therefore to be a viable alternative to the existing solutions not only for service robotics but possibly also in industry

This work presents the modeling of PZT (Lead Zirconate Titanate) intended for using in low frequency mechanical movement applications such as prosthetic legs. This includes the simplified PZT electromechanical modeling based on PSCAD/PSPICE, simulation and experiment. The model can emulate the behavior of the PZT in variety conditions. The simulation and experimental results well agree with each other. The benefits of the model are the easiness of analyzing and studying the behavior of PZT when the conditions or applications of use are changed such as in the case of using full-bridge diode rectifier, buck/boost converter, bridgeless rectifier, and series or parallel PZT modules.

This paper presents a reconfigurable spherical robot. The reconfigurable spherical robot can be reconfigured into a form of two interconnected hemispheres with three legs equipped with three omni-directional wheels. A stable reconfiguration control algorithm is constructed to change the robot from spherical shape to two halves of interconnected hemispheres and three legged-wheeled expansions. This work also constructs a transformation controller for the robot which uses an accelerometer to sense its orientation. The performance analysis shows that our reconfigurable robot prototype can transform from spherical shape (dormant mode) into two inter connected hemispheres where the three leg-wheels are projected out of the shells (transformed mode) and vice versa. After the transformation into the three leg-wheel configuration, the robot can autonomously move in L-shaped and U-shaped areas as well as narrowing pathways.

This paper presents a novel method for trajectory generation based on dynamic movement primitives DMPs treated from a control theoretical perspective. We extended the key ideas from the original DMP formalism by introducing a velocity convergence mechanism in the reformulated system. Theoretical proof is given to guarantee its validity. The new method can deal with complex paths as a whole. Based on this, we can generate smooth trajectories with automatically generated transition zones, satisfy position- and velocity boundary conditions at start and endpoint with high precision, and support multiple via-point applications. Theoretic proof of this method and experiments are presented

Nature apparently has succeeded in evolving biomechanics and creating neural mechanisms that allow living systems like walking animals to perform various sophisticated behaviors, e.g. and different gaits, climbing, turning, orienting, obstacle avoidance, attraction, anticipation. This shows that general principles of nature can provide biological inspiration for robotic designs or give useful hints of what is possible and design ideas that may have escaped our consideration. Instead of starting from scratch, this article presents how the biological principles can be used for mechanical design and control of walking robots, in order to approach living creatures in their level of performance. Employing this strategy allows us to successfully develop versatile, adaptive, and autonomous walking robots. Versatility in this sense means a variety of reactive behaviors including memory guidance, while adaptivity implies online learning capabilities. Autonomy is an ability to function without continuous human guidance. These three key elements are achieved under modular neural control and learning. In addition, the presented neural control technique is shown to be a powerful method of solving sensor-motor coordination problems of high complexity systems

Biomechanical studies of human walking reveal that compliance plays an important role at least in natural and smooth motions as well as for self-stabilization. Inspired by this, we present here the development of a new lower leg segment of the dynamic biped robot "RunBot". This new lower leg segment features a compliant ankle connected to a flat foot. It is mainly employed to realize robust self-stabilization in a passive manner. In general, such self-stabilization is achieved through mechanical feedback due to elasticity. Using real-time walking experiments, this study shows that the new lower leg segment improves dynamic walking behavior of the robot in two main respects compared to an old lower leg segment consisting of rigid ankle and curved foot: 1) it provides better self-stabilization after stumbling and 2) it increases passive dynamics during some stages of the gait cycle of the robot i.e., when the whole robot moves unactuated. As a consequence, a combination of compliance (i.e., the new lower leg segment) and active components (i.e., actuated hip and knee joints) driven by a neural mechanism (i.e., reflexive neural control) enables RunBot to perform robust self stabilization and at the same time natural, smooth, and energy efficient walking behavior without high control effort.

This paper presents the control system for our novel hyper-redundant chain robot HRCR system 3D-Trunk demonstrating an operational principle which is much different from traditional solutions. Its main features are that all the joints are passive, state controllable and share common inputs introduced by wire-driven control. For this unique design, a force-oriented method is employed to control the driving wires. The mechanical analysis, as well as an analysis of the differential driven mechanism of this design is formulated. The design of a novel wire tension state sensing component and its operation are also described. The system is controlled by distributed embedded controllers. The actuators coordination mechanism and the Bang-Bang controller-based closed-loop control implementation of this novel prototype are discussed from a mechatronic system level. Thus, this paper, together with a predecessor 1, presents all required details allowing for building and controlling 3D-Trunk

Critical behavior in neural networks characterized by scale-free event distributions and brought about by self-regulatory mechanisms such as short-term synaptic dynamics or homeostatic plasticity, is believed to optimize sensitivity to input and information transfer in the system. Although theoretical predictions of the spike distributions have been confirmed by in-vitro experiments, in-vivo data yield a more complex picture which might be due to the in-homogeneity of the network structure, leakage in currents or massive driving inputs which has so far not been comprehensively covered by analytical or numerical studies.

In this paper, the characteristic of the piezoelectric material, Macro Fiber Composites (MFC), has been investigated by comparison between the electrical equivalent circuit based simulation and the experimental result. The operational factors such as internal impedance and frequency which affect the maximum power output of the piezoelectric are systematically determined. The effect from the characteristic of the capacity after the rectifier circuit of the energy harvesting circuit in order to achieve the suitable energy storage is also mentioned. Some basic characteristic are tested and measured based on standard energy harvest kit and commercial MFC.

Execution of a manipulation after learning from demonstration many times requires intricate planning and control systems or some form of manual guidance for a robot. Here we present a framework for manipulation execution based on the so called "Semantic Event Chain" which is an abstract description of relations between the objects in the scene. It captures the change of those relations during a manipulation and thereby provides the decisive temporal anchor points by which a manipulation is critically defined. Using semantic event chains a model of a manipulation can be learned. We will show that it is possible to add the required control parameters (the spatial anchor points) to this model, which can then be executed by a robot in a fully autonomous way. The process of learning and execution of semantic event chains is explained using a box pushing example.

Generalizing objects in an action-context by a robot, for example addressing the problem: "Which items can be cut with which tools?", is an unresolved and difficult problem. Answering such a question defines a complete action class and robots cannot do this so far. We use a bootstrapping mechanism similar to that known from human language acquisition, and combine languagewith image-analysis to create action classes built around the verb (action) in an utterance. A human teaches the robot a certain sentence, for example: "Cut a sausage with a knife", from where on the machine generalizes the arguments (nouns) that the verb takes and searches for possible alternative nouns. Then, by ways of an internet-based image search and a classification algorithm, image classes for the alternative nouns are extracted, by which a large "picture book" of the possible objects involved in an action is created. This concludes the generalization step. Using the same classifier, the machine can now also perform a recognition procedure. Without having seen the objects before, it can analyze a visual scene, discovering, for example, a cucumber and a mandolin, which match to the earlier found nouns allowing it to suggest actions like: "I could cut a cucumber with a mandolin". The algorithm for generalizing objects by analyzing/anguage (GOAL) presented here, allows, thus, generalization and recognition of objects in an action-context. It can then be combined with methods for action execution (e.g. action generation-based on human demonstration) to execute so far unknown actions.

It has been shown that plasticity is not a fixed property but, in fact, changes depending on the location of the synapse on the neuron and/or changes of biophysical parameters. Here we investigate how plasticity is shaped by feedback inhibition in a cortical microcircuit. We use a differential Hebbian learning rule to model spike-timing dependent plasticity and show analytically that the feedback inhibition shortens the time window for LTD during spike-timing dependent plasticity but not for LTP. We then use a realistic GENESIS model to test two hypothesis about interneuron hypofunction and conclude that a reduction in GAD67 is the most likely candidate as the cause for hypofrontality as observed in Schizophrenia

Human environments are challenging for robots, which need to be trainable by lay people and learn new behaviours rapidly without disrupting much the ongoing activity. A system that integrates AI techniques for planning and learning is here proposed to satisfy these strong demands. The approach rapidly learns planning operators from few action experiences using a competitive strategy where many alternatives of cause-effect explanations are evaluated in parallel, and the most successful ones are used to generate the operators. The success of a cause-effect explanation is evaluated by a probabilistic estimate that compensates the lack of experience, producing more confident estimations and speeding up the learning in relation to other known estimates. The system operates without task interruption by integrating in the planning-learning loop a human teacher that supports the planner in making decisions. All the mechanisms are integrated and synchronized in the robot using a general decision-making framework. The feasibility and scalability of the architecture are evaluated in two different robot platforms: a Stäubli arm, and the humanoid ARMAR III.

This paper presents a probabilistic algorithm for lane shape estimation in an urban environment which is important for example for driver assistance systems and autonomous driving. For the first time, we bring together the so-called Partitioned Particle filter, an improvement of the traditional Particle filter, and the linear-parabolic lane model which alleviates many shortcomings of traditional lane models. The former improves the traditional Particle filter by subdividing the whole state space of particles into several subspaces and estimating those subspaces in a hierarchical structure, such that the number of particles for each subspace is flexible and the robustness of the whole system is increased. Furthermore, we introduce a new statistical observation model, an important part of the Particle filter, where we use multi- kernel density to model the probability distribution of lane parameters. Our observation model considers not only color and position information as image cues, but also the image gradient. Our experimental results illustrate the robustness and efficiency of our algorithm even when confronted with challenging scenes

Recognizing manipulations performed by a human and the transfer and execution of this by a robot is a difficult problem. We address this in the current study by introducing a novel representation of the relations between objects at decisive time points during a manipulation. Thereby, we encode the essential changes in a visual scenery in a condensed way such that a robot can recognize and learn a manipulation without prior object knowledge. To achieve this we continuously track image segments in the video and construct a dynamic graph sequence. Topological transitions of those graphs occur whenever a spatial relation between some segments has changed in a discontinuous way and these moments are stored in a transition matrix called the semantic event chain (SEC). We demonstrate that these time points are highly descriptive for distinguishing between different manipulations. Employing simple sub-string search algorithms, SECs can be compared and type-similar manipulations can be recognized with high confidence. As the approach is generic, statistical learning can be used to find the archetypal SEC of a given manipulation class. The performance of the algorithm is demonstrated on a set of real videos showing hands manipulating various objects and performing different actions. In experiments with a robotic arm, we show that the SEC can be learned by observing human manipulations, transferred to a new scenario, and then reproduced by the machine.

When describing robot motion with dynamic motion primitives DMPs, goal-trajectory endpoint, shape and temporal scaling parameters are used. In reinforcement learning with DMPs, usually goals and temporal scaling parameters are pre-defined and only the weights for shaping a DMP are learned. Many tasks, however, exist where the best goal position is not a priori known, requiring to learn it. Thus, here we specifically address the question of how to simultaneously combine goal and shape parameter learning. This is a difficult problem because both parameters could easily interfere in a destructive way. We apply value function approximation techniques for goal learning and policy gradient methods for shape learning. Specifically, we use policy improvement with path integrals and natural actor-critic for the policy gradient approach. Methods are analyzed with simulations and implemented on a real robot setup. Results for learning from scratch, learning initialized by human demonstration, as well as for modifying the tool for the learned DMPs are presented. We observe that the combination goal- together with shape learning is stable and robust within large parameter regimes. Learning converges quickly even in the presence of large disturbances, which makes this combined method suitable for robotic applications

The generation of complex movement patterns, in particular in cases where one needs to smoothly and accurately join trajectories, is still a difficult problem in robotics. This paper presents a novel approach for joining of several dynamic movement primitives DMPs based on a modification of the original formulation for DMPs. The new method produces smooth and natural transitions in position as well as velocity space. The properties of the method are demonstrated by applying it to simulated handwriting generation implemented on a robot, where an adaptive algorithm is used to learn trajectories from human demonstration. These results demonstrate that the new method is a feasible alternative for trajectory learning and generation and its accuracy and modular character has potential for various robotics applications

n this contribution, we introduce a new state estimation filter for nonlinear estimation and sensor fusion, which we call cen- tral difference information filter CDIF. As we know, the ex- tended information filter EIF has two shortcomings: one is the limited accuracy of the Taylor series linearization method, the other is the calculation of the Jacobians. These shortcom- ings can be compensated by utilizing sigma point information filters SPIFs, e.g. and the unscented information filter UIF, which uses deterministic sigma points to approximate the distribution of Gaussian random variables and does not require the calculation of Jacobians. As an alternative to the UIF, the CDIF is derived by using Stirlings interpolation to generate sigma points in the SPIFs architecture, which uses less parameters, has lower computational cost and achieves the same accuracy as UIF. To demonstrate the performance of our al- gorithm, a classic space vehicle reentry tracking simulation is used

Autonomous cognitive robots must be able to interact with the world and reason about their interactions. On the one hand, physical interactions are inherently continuous, noisy, and require feedback. On the other hand, the knowledge needed for reasoning about high-level objectives and plans is more conveniently expressed as symbolic predictions about state changes. Bridging this gap between control knowledge and abstract reasoning has been a fundamental concern of autonomous robotics. This paper proposes a formalism called an Object-Action Complex as the basis for symbolic representations of sensorimotor experience. OACs are designed to capture the interaction between objects and associated actions in artificial cognitive systems. This paper defines a formalism for describing object action relations and their use for autonomous cognitive robots, and describes how OACs can be learned. We also demonstrate how OACs interact across different levels of abstraction in the context of two tasks: the grounding of objects and grasping accordances, and the execution of plans using grounded representations

This works present the design of an energy harvesting system using smart materials for self power generation of upper and lower prosthetic legs. The smart materials like Piezo-Composites, Piezo Flexible Film, Macro Fiber Composites, and PZT have been employed and modified to be appropriately embedded in the prosthesis. The movements of the prosthesis would extract and transfer energy directly from the piezoelectric via a converter to a power management system. Afterward, the power management system manages and accumulates the generated electrical energy to be sufficient for later powering electronic components of the prosthesis. Here we show our preliminary experimental results of energy harvesting and efficiency in peak piezoelectric voltages during step up and continuous walking for a period of time.

Efficient segmentation of color images is important for many applications in computer vision. Non-parametric solutions are required in situations where little or no prior knowledge about the data is available. In this paper, we present a novel parallel image segmentation algorithm which segments images in real-time in a non-parametric way. The algorithm finds the equilibrium states of a Potts model in the superparamagnetic phase of the system. Our method maps perfectly onto the Graphics Processing Unit GPU architecture and has been implemented using the framework NVIDIA Compute Unified Device Architecture CUDA. For images of 256 x 320 pixels we obtained a frame rate of 30 Hz that demonstrates the applicability of the algorithm to video-processing tasks in real-time1

The sigma-point information filters employ a number of deterministic sigma-points to calculate the mean and covariance of a random variable which undergoes a nonlinear transformation. These sigma-points can be generated by the unscented transform or Stirlings interpolation, which corresponds to the unscented information filter (UIF) and the central difference information filter (CDIF) respectively. In this technical note, we develop the square-root extensions of UIF and CDIF, which have better numerical properties than the original versions, e.g., improved numerical accuracy, double order precision and preservation of symmetry. We also show that the square-root unscented information filter (SRUIF) might lose the positive-definiteness due to the negative Cholesky update, whereas the square-root central difference information filter (SRCDIF) has only positive Cholesky update. Therefore, the SRCDIF is preferable to the SRUIF concerning the numerical stability.

Synaptic scaling is a slow process that modifies synapses, keeping the firing rate of neural circuits in specific regimes. Together with other processes, such as conventional synaptic plasticity in the form of long term depression and potentiation, synaptic scaling changes the synaptic patterns in a network, ensuring diverse, functionally relevant, stable, and input-dependent connectivity. How synaptic patterns are generated and stabilized, however, is largely unknown. Here we formally describe and analyze synaptic scaling based on results from experimental studies and demonstrate that the combination of different conventional plasticity mechanisms and synaptic scaling provides a powerful general framework for regulating network connectivity. In addition, we design several simple models that reproduce experimentally observed synaptic distributions as well as the observed synaptic modifications during sustained activity changes. These models predict that the combination of plasticity with scaling generates globally stable, input-controlled synaptic patterns, also in recurrent networks. Thus, in combination with other forms of plasticity, synaptic scaling can robustly yield neuronal circuits with high synaptic diversity, which potentially enables robust dynamic storage of complex activation patterns. This mechanism is even more pronounced when considering networks with a realistic degree of inhibition. Synaptic scaling combined with plasticity could thus be the basis for learning structured behavior even in initially random networks

We present a system that learns basic vision based driving skills from a human teacher. In contrast to much other work in this area which is based on simulation, or data obtained from simulation, our system is implemented as a multi-threaded, parallel CPU/GPU architecture in a real car and trained with real driving data to generate steering and acceleration control for road following. In addition it uses a novel algorithm for detecting independently moving objects IMOs for spotting obstacles. Both, learning and IMO detection algorithms, are data driven and thus improve above the limitations of model based approaches. The systems ability to imitate the teachers behavior is analyzed on known and unknown streets and the results suggest its use for steering assistance but limit the use of the acceleration signal to curve negotiation. We propose that this ability to adapt to the driver has high potential for future intelligent driver assistance systems since it can serve to increase the drivers security as well as the comfort, an important sales argument in the car industry

In this paper, we present energy harvesting systems in a prosthetic leg using piezoceramic Macro Fiber Composites MFCs and their models using artificial neural networks. The piezoceramic MFCs are implemented at the sole and heel of the leg and transform impact forces into electrical power during walking. The neural model of the energy harvesting system installed at the sole is developed on the basis of a standard feedforward backpropagation neural network. On the other hand, the neural model of the energy harvesting system installed at the heel is manually synthesized from different neural modules networks. Experimental results show that these neural models can appropriately transform the impact forces detected by force sensing resistors FSRs into the electrical responses of the piezoceramic MFCs. The models will be used to study and analyze dynamical behaviors of the piezoelectric materials with respect to walking