The robot-based tracking of highly dynamic end point motions of deformable linear objects (DLO) remains challenging due to its non-linear behavior. Since simple feedback control is infeasible, model-based control offers potential to account for the non-linear effects, but requires computation efficient and accurate models. Promising results have been achieved utilizing data-driven models that introduce a latent kinematic chain as model of the DLO and mapping measurements of the tip position in its latent joint space, in which the dynamic motion model is learned. So far, this approach has the limitation that it can not handle situations of incomplete sensory information, for instance if occlusion occurs. Consequently, this paper introduces a fusion network architecture capable of making predictions even if sensory information is incomplete. We achieve additional state estimation of the latent joint state by learning a data driven inverse kinematics with help of wrench measurements at the DLO base and evaluate our approach by simulating occlusion. We demonstrate the computational effectiveness of our approach for in the loop control tasks.

Concentric tube continuum robots (CTCRs) belong to the family of continuum robots with applications in minimally invasive surgeries. Because of this application domain, measuring the external forces along the body of the robot is paramount. CTCRs are made up of thin elastic rods and are intended to be applied inside the human body, where conventional sensor-based measurements are not feasible. Consequently, research is resorting to estimate the forces through geometric, numeric, or optimization methods. However, these methods often suffer from slow convergence. In this paper, we introduce a novel data-driven approach for estimating contact forces along the body of a CTCR that offers an estimation precision comparable to the current state-of-the-art optimization-based approaches, but exhibits nearly two orders of magnitude faster convergence. The proposed method is scalable and exhibits a significant performance in response to a wide range of external forces. The approach was evaluated in simulations and on a real 2-tube CTCR.

Early detection of laryngeal tumors is critical for their successful therapy.In this paper, we investigate how hyperspectral (HS) imaging can contribute to this aim based on an in-vivo data set of 13 HS image cubes recorded in clinical practice.We perform semantic segmentation with a tailored U-Net trained on labels provided by the clinicians.We specifically investigate the influence of exposure time during image acquisition, the suitable wavelengths to determine the most informative image channels, and present quantitative results on accuracy and the AUC measure.* The data set collection was funded by the German Cancer Aid within the project framework "Early detection of laryngeal cancer by means of Hyperspectral Imaging (109825110275)".

Numerous robotic and control applications have strict real-time requirements, which, when violated, result in reduced quality of service or, in case of safety critical applications, might even have catastrophic consequences. To ensure that certain real-time constraints are satisfied, roboticists have relied on real-time safe frameworks, environments and middleware. With the introduction of ROS 2, alongside kernel patches such as PREEMPT_RT, there is an abundance of solutions to pick from. This paper compares OROCOS and ROS 2 over PREEMPT_RT and vanilla Linux kernels in a variety of benchmarks and draws conclusions on their performance in real-time critical applications. The outcome of the benchmark shows comparable performances under normal conditions. However, when the system is under stress both frameworks suffer in different fashions. Furthermore, the results show an accumulating error which over time violates the real-time requirements in both frameworks. These findings are paramount in conducting real world application with real-time constraints.

We propose a novel hierarchical online learning scheme for fast and efficient bootstrapping of sensorimotor skills. Our scheme permits rapid data-driven robot model learning in a "learning while behaving" fashion. It is updated continuously to adapt to time-dependent changes and driven by an intrinsic motivation signal. It utilizes an online associative radial basis function network, which is the first associative dynamic network to be constructed from scratch with high stability. Moreover, we propose a parameter-sharing technique to increase efficiency, stabilize the online scheme, avoid exhaustive parameter tuning, and speed up the learning process. We apply our proposed algorithms on a 7-DoF physical robot manipulator and demonstrate their performance and efficiency.

Tracking multiple prioritized tasks simultaneously with redundant robots have been investigated extensively over the last decades. Recent research focuses on combining advantages from both classical soft and strict prioritization schemes which is non-trivial. Among the proposed methods to tackle this issue, Generalized Hierarchical Control (GHC) seems to have a reasonable performance, however, it does not include a weighting matrix in the computation of the nullspace projection operator and hence cannot construct dynamically-consistent stack-of-tasks hierarchies as a special case. We extend GHC by adding dynamic-consistency to the control scheme and refer to it as DynGHC. The extension is also advantageous when choosing non-strict priorities because inertia coupling between tasks is reduced. DynGHC allows to smoothly rearrange priorities which is important for robots acting in dynamically changing contexts. Comparative simulations with a 4 DOF planar manipulator and a KUKA LWR validate our approach. Matlab and C++ source code is made available.

In this paper, we look at the challenge of human robot interaction in locomotion. We consider a hand-in-hand interaction scenario where a human compliantly interacts with the upper-body of an impedance controlled humanoid. By exploring the velocity transmission of the robot arms, and the interaction in terms of robot arms manipulation quality evaluated through the monitoring of their manipulability the proposed method derives suitable reactive steps in appropriate directions to ensure that the robot manipulation ability is maintained with the robot arms providing high capacity of motion along the different directions. The proposed approach can be combined with different walking pattern generators and is not tailored to a specific one used in this work. The results of the proposed method are experimentally validated on the COMAN + humanoid robot showing the efficacy of the method to generate reactive stepping driven by the interaction and manipulation motion of the human operator. Besides, the work also provides a real-time software architecture to control humanoid COMAN+, but it is also flexible to be used for the control of other robot platforms.

We explore the characteristics of secondary contacts when applying forces with the end-effector of a robot and address the question when these secondary contacts can increase maximum applicable end-effector forces or reduce required actuator efforts. To this end, we formalize the effect of such secondary contacts in terms of required actuator efforts and derive efficiency bounds depending on the contact characteristics and robot configuration. Our findings are confirmed by experiments with a redundant serial manipulator.

The plasticity of the human nervous system allows us to acquire an open-ended repository of sensorimotor skills in adulthood, such as the mastery of tools, musical instruments or sports. How novel sensorimotor skills are learned from scratch is yet largely unknown. In particular, the so-called inverse mapping from goal states to motor states is underdetermined because a goal can often be achieved by many different movements (motor redundancy). How humans learn to resolve motor redundancy and by which principles they explore high-dimensional motor spaces has hardly been investigated. To study this question, we trained human participants in an unfamiliar and redundant visually-guided manual control task. We qualitatively compare the experimental results with simulation results from a population of artificial agents that learned the same task by Goal Babbling, which is an inverse-model learning approach for robotics. In Goal Babbling, goal-related feedback guides motor exploration and thereby enables robots to learn an inverse model directly from scratch, without having to learn a forward model first. In the human experiment, we tested whether different initial conditions (starting positions of the hand) influence the acquisition of motor synergies, which we identified by Principal Component Analysis in the motor space. The results show that the human participants' solutions are spatially biased towards the different starting positions in motor space and are marked by a gradual co-learning of synergies and task success, similar to the dynamics of motor learning by Goal Babbling. However, there are also differences between human learning and the Goal Babbling simulations, as humans tend to predominantly use Degrees of Freedom that do not have a large effect on the hand position, whereas in Goal Babbling, Degrees of Freedom with a large effect on hand position are used predominantly. We conclude that humans use goal-related feedback to constrain motor exploration and resolve motor redundancy when learning a new sensorimotor mapping, but in a manner that differs from the current implementation of Goal Babbling due to different constraints on motor exploration.

We consider a virtual manipulator in grasping scenarios which allows us to capture the effect of the object dynamics. This modeling approach turns a multi-arm robot into an underactuated system. We observe that controlling floating-base multi-leg robots is fundamentally similar. The Projected Inverse Dynamics Control approach is employed for decoupling contact consistent motion generation and controlling contact wrenches. The proposed framework for underactuated robots has been evaluated on an enormous robot hand composed of four KUKA LWR IV+ representing fingers cooperatively manipulating a 9kg box with total 28 actuated DOF and six virtual DOF representing the object as additional free-floating robot link. Finally, we validate the same approach on ANYmal, a floating-base quadruped with 12 actuated DOF. Experiments are performed both in simulation and real world.

Projection operators are widely employed in multi-objective robot control. It is an open research question how to achieve continuous transitions between different idempotent projectors which is required for dynamic task priority rearrangement. We formalize projection shaping, providing a solution to deal with rank changes in a smooth fashion. Furthermore, we derive meaningful shaping operators and show that damped least squares is a special case of our general formulation. Finally, we extend the Stack-of-Tasks prioritization scheme for continuous priority rearrangement of single task dimensions. Simulation results validate our approach.

When learning forward and inverse kinematics maps of manipulators, usually little attention is paid to data-efficiency, i.e., the accuracy gained per action-outcome sample. This paper examines properties of popular (online) learning techniques and demonstrates that - regardless of the employed exploration strategy - the structure of kinematics mappings does not allow for a practically viable trade-off between the number of samples and the resulting approximation error for manipulators with more than a few DoFs - unless tailored parametric models are employed. We discuss suitable choices for these parametric models for both rigid and elastic discretely-actuated robots and compare their data -efficiency to that of popular exploratory learning approaches relying on non-parametric models. Our theoretical considerations are confirmed by various experimental results for inverse kinematics mappings of rigid and omnielastic manipulators.

Motor rehabilitation is in increasingly high demand to deal with minor functional motor impairments resulting from stroke, cerebellar ataxia, or Parkinson's disease. Juggling physiotherapy has shown to induce brain plasticity and to improve coordination and balance in this context. The physiotherapy, however, relies on large number of repetitions to be effective which prompts to deploy robots to release the burden on therapists both in terms of time as well as physical strain. This paper provides a framework to enable juggling games for patients in interacting with robots through Virtual Reality (VR). A set of throwing motions is recorded from the therapist and is retargeted to the humanoid robot COMAN's wrist. The respective whole-body motion is then solved in a stack of Quadratic Programs (QP) in a real-time architecture that integrates OROCOS and Gazebo. The resulting motion is finally streamed to VR for animation of the robot and the thrown ball, which the user can catch in VR using a controller device. We regard the VR setting as an essential step towards physiotherapeutic robotic juggling, because it ensures safety of the patients and effective testing of the methods and already has potential for actual therapeutic intervention. The control framework, however, is already validated in this paper for switching to full real-time operation on the physical robot.

Tactile surface sensors (TSSs) are often utilized for contact management in human robot interaction scenarios. To provide added value in these applications, online robot programming approaches may exploit TSSs as gesture input devices. To this end, we introduce an invariant, compact gesture representation which facilitates robust and efficient online gesture recognition even for very small training data sets. The proposed two-stage recognition approach permits reliable gesture classification as well as convenient parameter extraction in the presence of typical disturbances affecting TSSs attached to manipulator links. Our experimental results demonstrate the remarkable recognition performance of the proposed approach using a set of 16 gestures and data of up to 31 subjects.

This paper introduces a Personalized Adaptive Stiffness controller for physical Human-Robot Interaction and validates its performance in an extensive user study with 49 participants. The controller is calibrated to the user's force profile to account for inter-user variance and individual differences. The user study compares the new scheme to conventional fixed stiffness or gravitation compensation controllers on the 7-DOF KUKA LWR IVb by employing two typical joint-manipulation tasks. Somewhat surprisingly, the experiments suggest that for simpler tasks a standard fixed controller may perform sufficiently well and that respective task dependency strongly prevails over individual differences. In the more complex task, quantitative and qualitative results clearly show differences between the different control modes and a both a performance gains and a user preference for the Personalized Adaptive Stiffness controller.

This paper focuses on Human-Robot Interaction (HRI) in a scenario where a robot has to recover from an assembly problem. For this purpose, a human interacts with the robot so that the task-specific knowledge and experience of the human can be used for recovery. The question which interaction method is the most successful for this problem will be answered by a user study with 31 participants. Four different interaction methods are compared: space mouse, keyboard, kinesthetic guidance and a kinesthetic guidance with adaptive stiffness. The results clearly show that the kinesthetic guidance methods have superior performance and user satisfaction compared to the other considered methods.

In this paper, we show the generalization of an inverse dynamic model for KUKA LWR IV+ under load mass variations. We use a modular approach based on regression in the model space. First, inverse dynamic models for the known masses are learned using a recently proposed approach called Independent Joint Learning (IJL). In IJL the torque errors due to unmodeled dynamics of the real robot are estimated using only joint-local information. Second, a mapping from load mass to model parameters of torque error model is learned in order to generalize the inverse dynamics to new load masses. The modular approach improves the accuracy of an existing KUKA LWR IV+ inverse dynamic model. The results are compared with a single step IJL approach. The results show the excellent generalization for new load masses using regression in the model space.

“Happy New Year!” At the beginning of 2016, I would like to take this opportunity to wish everyone a very happy, healthy, and prosperous new year! It is my great honor and privilege to serve as the Editor-in-Chief (EiC) of the IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS (TNNLS), and I am excited to write this Editorial to start a new journey with you all.

Kinesthetic teaching is a promising approach to acquire robot skills in an intuitive way. This paper focuses on learning skills that do not solely rely on kinematics but also need to take into account interaction forces. We present three novel concepts towards learning such force interaction skills. Firstly, we determine segments from a small number of continuous kinesthetic demonstrations using contact information. Secondly, we associate each segment with a movement primitive, and determine its composition, i.e., the control variables and reference frames that allow to reproduce the demonstrated task. Lastly, we propose a concept to determine the transitions between the primitives during reproduction. The proposed methods are evaluated on a box pulling and flipping task, and show very good generalization abilities for objects with different geometries, and situations with different object arrangements.

This paper introduces a model-driven approach for engineering complex movement control architectures based on motion primitives, which in recent years have been a central development towards adaptive and flexible control of complex and compliant robots. We consider rich motor skills realized through the composition of motion primitives as our domain. In this domain we analyze the control architectures of representative example systems to identify common abstractions. It turns out that the introduced notion of motion primitives implemented as dynamical systems with machine learning capabilities, provide the computational building block for a large class of such control architectures. Building on the identified concepts, we introduce domain-specific languages that allow the compact specification of movement control architectures based on motion primitives and their coordination respectively. Using a proper tool chain, we show how to employ this model-driven approach in a case study for the real world example of automatic laundry grasping with the KUKA LWR-IV, where executable source-code is automatically generated from the domain-specific language specification.

Simultaneous mastering of multiple tasks during motion generation is challenging. Traditional null-space based approaches for redundant robots implement a strict, hierarchical prioritization for tracking multiple objectives. In consequence, these schemes are not suited to impose smooth priorities or changing them during motion execution. A recently developed mixture of controller approach superpose torques from several control modules for motion generation and thereby enables to flexibly impose priorities for pursuing different goals in parallel. The main contribution of this paper is the development of a framework which allows for automatic derivation of suitable mixture coefficients which represent priorities. The functionality of the optimization framework is demonstrated for a virtual 3 DOF pendulum and the humanoid robot COMAN.

Bionic soft robots offer exciting perspectives for more flexible and safe physical interaction with the world and humans. Unfortunately, their hardware design often prevents analytical modeling, which in turn is a prerequisite to apply classical automatic control approaches. On the other hand, also modeling by means of learning is hardly feasible due to many degrees of freedom, high-dimensional state spaces and the softness properties like e.g. mechanical elasticity, which cause limited repeatability and complex dynamics. Nevertheless, the realization of basic control modes is important to leverage the potential of soft robots for applications. We therefore propose a hybrid approach combining classical and learning elements for the realization of an interactive control mode for an elastic bionic robot. It superimposes a low-gain feedback control with a feed-forward control based on a learned simplified model of the inverse dynamics which considers only equilibria of the robot's dynamics. We demonstrate on the Bionic Handling Assistant how a respective inverse equilibrium model can be learned and effectively exploited for quick and agile control. In a second step, the control scheme is extended to an active compliant control mode. It implements a kind of gravitation compensation to allow for kinesthetic teaching of the robot based on the implicit knowledge of gravitational and mechanical forces that are encoded in the learned equilibrium model. We finally discuss that this control scheme may be implemented also on other soft robots to provide the avenue towards their applications in general manipulation tasks.

Motion primitives are an established paradigm to generate complex motions from simpler building blocks. A much less addressed issue is at which level to encode and organize a library of motion primitives, and how to retrieve motion primitives from a library that fit a particular task. This paper proposes a parameterized skill memory, which organizes a set of motion primitives in a low-dimensional, topology-preserving embedding space. The skill memory acts as a pivotal mechanism that links low-dimensional skill parametrizations to motion primitive parameters and complete motion trajectories. The skill memory is implemented by means of a dynamical system which features continuous generalization of motion shapes. It is shown that the low-dimensional skill parametrization is beneficial for efficient, reward-based retrieval of motion primitives and simplifies the shaping of reward functions. The excellent generalization of motion shapes by parameterized skill memories from few training examples is demonstrated in a bimanual manipulation task with the humanoid robot iCub.

The paper addresses path planning for a redundant robot arm that is maneuvering in confined spaces, where neither an explicit model nor external perception of the possibly frequently changing environment is available. Our approach is rather solely based on data from kinesthetic demonstrations of feasible configurations provided by a user. The key challenge is to create a graph-based representation of the demonstrated free space incrementally and online by means of an specifically tailored instantaneous topological map at runtime. Subsequent application of standard graph-based planning in combination with a learned generalization of the demonstrated redundancy resolution then enables the robot to safely move in the realm of the demonstrated task space areas. This model-free approach greatly enhances configurability and flexibility of the robot for assistance applications, where movement capabilities need to be realized without explicit programming.

We present an approach to learn the inverse kinematics of the “bionic handling assistant”-an elephant trunk robot. This task comprises substantial challenges including high dimensionality, restrictive and unknown actuation ranges, and nonstationary system behavior. We use a recent exploration scheme, online goal babbling, which deals with these challenges by bootstrapping and adapting the inverse kinematics on the fly. We show the success of the method in extensive real-world experiments on the nonstationary robot, including a novel combination of learning and traditional feedback control. Simulations further investigate the impact of nonstationary actuation ranges, drifting sensors, and morphological changes. The experiments provide the first substantial quantitative real-world evidence for the success of goal-directed bootstrapping schemes, moreover with the challenge of nonstationary system behavior. We thereby provide the first functioning control concept for this challenging robot platform.

Facilitating efficient programming-by-demonstration methods for advanced robot systems is an ongoing research challenge. This paper addresses one important challenge in this area, which is the programming of kinematically redundant robots. We argue that standard programming-by-demonstration methods for teaching task-space trajectories on a redundant robot using physical human-robot interaction are too complex for non-expert human tutors. We therefore introduce a new interaction and control concept for redundant robot systems, Assisted Gravity Compensation, based on a hierarchical control scheme, separating task-space programming from the redundancy resolution. The user is actively assisted by a given redundancy resolution while kinesthetically teaching task-space trajectories. This control scheme is implemented on our experimental robot system called FlexIRob and we briefly present results of a kinesthetic teaching experiment obtained in a larger field study on physical Human-Robot Interaction with 48 industrial workers. These results show, that the Assisted Gravity Compensation reduces the complexity of a kinesthetic teaching task, which is revealed by an improved task performance, making kinesthetic teaching an efficient programming-by-demonstration method for redundant robots.

Nonlinear dynamical systems are a promising representation to learn complex robot movements. Besides their undoubted modeling power, it is of major importance that such systems work in a stable manner. We therefore present a neural learning scheme that estimates stable dynamical systems from demonstrations based on a two-stage process: first, a data-driven Lyapunov function candidate is estimated. Second, stability is incorporated by means of a novel method to respect local constraints in the neural learning. We show in two experiments that this method is capable of learning stable dynamics while simultaneously sustaining the accuracy of the estimate and robustly generates complex movements.

In the course of trial-and-error learning, the results of actions, manifested as rewards or punishments, occur often seconds after the actions that caused them. How can a reward be associated with an earlier action when the neural activity that caused that action is no longer present in the network? This problem is referred to as the distal reward problem. A recent computational study proposes a solution using modulated plasticity with spiking neurons and argues that precise firing patterns in the millisecond range are essential for such a solution. In contrast, the study reported in this letter shows that it is the rarity of correlating neural activity, and not the spike timing, that allows the network to solve the distal reward problem. In this study, rare correlations are detected in a standard rate-based computational model by means of a threshold-augmented Hebbian rule. The novel modulated plasticity rule allows a randomly connected network to learn in classical and instrumental conditioning scenarios with delayed rewards. The rarity of correlations is shown to be a pivotal factor in the learning and in handling various delays of the reward. This study additionally suggests the hypothesis that short-term synaptic plasticity may implement eligibility traces and thereby serve as a selection mechanism in promoting candidate synapses for long-term storage.

The recent advent of compliant and kinematically redundant robots poses new research challenges for human-robot interaction. While these robots provide a great degree of flexibility for the realization of complex applications, the flexibility gained generates the need for additional modeling steps and definition of criteria for redundancy resolution constraining the robot's movement generation. The explicit modeling of such criteria usually require experts to adapt the robot's movement generation subsystem. A typical way of dealing with this configuration challenge is to utilize kinesthetic teaching by guiding the robot to implicitly model the specific constraints in task and configuration space. We argue that current programming-by-demonstration approaches are not efficient for kinesthetic teaching of redundant robots and show that typical teach-in procedures are too complex for novice users. In order to enable non-experts to master the configuration and programming of a redundant robot in the presence of non-trivial constraints such as confined spaces, we propose a new interaction scheme combining kinesthetic teaching and learning within an integrated system architecture. We evaluated this approach in a user study with 49 industrial workers at HARTING, a medium-sized manufacturing company. The results show that the interaction concepts implemented on a KUKA Lightweight Robot IV are easy to handle for novice users, demonstrate the feasibility of kinesthetic teaching for implicit constraint modeling in configuration space, and yield significantly improved performance for the teach-in of trajectories in task space.

A major goal of current robotics research is to enable robots to become co-workers that collaborate with humans efficiently and adapt to changing environments or workflows. We present an approach utilizing the physical interaction capabilities of compliant robots with data-driven and model-free learning in a coherent system in order to make fast reconfiguration of redundant robots feasible. Users with no particular robotics knowledge can perform this task in physical interaction with the compliant robot, for example to reconfigure a work cell due to changes in the environment. For fast and efficient learning of the respective null-space constraints, a reservoir neural network is employed. It is embedded in the motion controller of the system, hence allowing for execution of arbitrary motions in task space. We describe the training, exploration and the control architecture of the systems as well as present an evaluation on the KUKA Light-Weight Robot. Our results show that the learned model solves the redundancy resolution problem under the given constraints with sufficient accuracy and generalizes to generate valid joint-space trajectories even in untrained areas of the workspace.

We present a dynamical system approach to learning forward and inverse kinematics of a humanoid upper body in associative radial basis function networks. Coupling of arm kinematics via the torso joints is modeled by dynamically coupling two networks learning the direct inverse kinematics of both torso-arm chains separately. Dividing the upper body kinematics in two problems significantly reduces the number of samples required for learning. Redundancies of the inverse kinematics are represented by multi-stable dynamics of the associative networks and are resolved dynamically depending on the current system state. The model is exploited for task space tracking in a feedback control framework.

We evaluate the use of continuum kinematics with constant curvature as a kinematic model for Festo's “Bionic Handling Assistant” (BHA). We introduce a new, elegant, and parameterless method to deal with geometric singularities in stretched positions, which allows to capture pure elongations that are not naturally expressed by the toroidal deformations underlying the constant curvature assumption. The stability of the method is shown with numeric simulations. We evaluate how well this model describes the BHA by using real-world position measurements as quantitative ground truth and find a good match between model and real BHA, with only 1% relative error. The model provides a practical, and highly efficient tool for the simulation and experimentation with continuum robots and is available as free software library1.

This paper presents an architecture for learning and reproducing movements with a robot in interaction with a human teacher. We focus on the movement representation and propose three enhancements to increase generalization capabilities: Firstly, we introduce a flexible task-level movement representation that is based on neuropsychological findings. Movement is represented in task-oriented frames of reference, and generalizes to a variety of different situations. Secondly, we propose a mechanism to decouple the task descriptors from the perceived objects in the robot's environment. This allows to formulate a set of generic controllers, and to interactively create associations with perceived objects. Thirdly, we introduce a method to dynamically modify the system's body schema to account for structural changes such as having grasped a tool. The changes are consistently treated in the kinematics computations. This permits to generalize movements to be carried out in different ways, for instance with different hands or bi-manually. A set of experiments in an interactive imitation learning situation underline the capabilities of the proposed concepts.

In this paper we present a new robot control and learning framework. By integrating previously presented as well as new methods, the robot is able to learn an invariant and generic movement representation from a human tutor. We argue that in order to apply such generic representations to new situations and thus create a flexible system, the use of interaction is beneficial. The interaction is based on a kinematically controlled model of a human tutor, which is used as a model-based filter and also for recognizing postures that influence the interaction. In addition, a new movement segmentation scheme is presented that is based on correlating movements by the tutor's hand with the salient objects in the scene. The focus of this paper is on the interactive learning aspects of the system and particular emphasis is given to an experiment in which the humanoid robot ASIMO learns from a human tutor. The system includes extensive generalization capabilities that result from an online adaption of the robot's body schema and the exploitation of inter-trial variance from multiple demonstrations. This enables the robot to reproduce the movement in new situations. For example, a stacking task that the tutor performed one-handed can be executed bi-manually by the robot.

Recent advances in the field of humanoid robotics increase the complexity of the tasks that such robots can perform. This makes it increasingly difficult and inconvenient to program these tasks manually. Furthermore, humanoid robots, in contrast to industrial robots, should in the distant future behave within a social environment. Therefore, it must be possible to extend the robot's abilities in an easy and natural way. To address these requirements, this work investigates the topic of imitation learning of motor skills. The focus lies on providing a humanoid robot with the ability to learn new bi-manual tasks through the observation of object trajectories. For this, an imitation learning framework is presented, which allows the robot to learn the important elements of an observed movement task by application of probabilistic encoding with Gaussian Mixture Models. The learned information is used to initialize an attractor-based movement generation algorithm that optimizes the reproduced movement towards the fulfillment of additional criteria, such as collision avoidance. Experiments performed with the humanoid robot ASIMO show that the proposed system is suitable for transferring information from a human demonstrator to the robot. These results provide a good starting point for more complex and interactive learning tasks.

Previous work [1] shows that the movement representation in task spaces offers many advantages for learning object-related and goal-directed movement tasks through imitation. It allows to reduce the dimensionality of the data that is learned and simplifies the correspondence problem that results from different kinematic structures of teacher and robot. Further, the task space representation provides a first generalization, for example wrt. differing absolute positions, if bi-manual movements are represented in relation to each other. Although task spaces are widely used, even if they are not mentioned explicitly, they are mostly defined a priori. This work is a step towards an automatic selection of task spaces. Observed movements are mapped into a pool of possibly even conflicting task spaces and we present methods that analyze this task space pool in order to acquire task space descriptors that match the observation best. As statistical measures cannot explain importance for all kinds of movements, the presented selection scheme incorporates additional criteria such as an attention-based measure. Further, we introduce methods that make a significant step from purely statistically-driven task space selection towards model-based movement analysis using a simulation of a complex human model. Effort and discomfort of the human teacher is being analyzed and used as a hint for important task elements. All methods are validated with real-world data, gathered using color tracking with a stereo vision system and a VICON motion capturing system.

We present a strategy for grasping of real world objects with two anthropomorphic hands, the three-fingered 9- DOF hydraulic TUM and the very dextrous 20-DOF pneumatic Bielefeld Shadow Hand. Our approach to grasping is based on a reach-pre-grasp-grasp scheme loosely motivated by human grasping. We comparatively describe the two robot setups, the control schemes, and the grasp type determination. We show that the grasp strategy can robustly cope with inaccurate control and object variation. We demonstrate that it can be ported among platforms with minor modifications. Grasping success is evaluated by comparative experiments performing a benchmark test on 21 everyday objects.

We present a biologically motivated architecture for object recognition that is capable of online learning of several objects based on interaction with a human teacher. The system combines biological principles such as appearance-based representation in topographical feature detection hierarchies and context-driven transfer between different levels of object memory. Training can be performed in an unconstrained environment by presenting objects in front of a stereo camera system and labeling them by speech input. The learning is fully online and thus avoids an artificial separation of the interaction into training and test phases. We demonstrate the performance on a challenging ensemble of 50 objects.

We present an on-line, robust, and efficient path planner for the redundant Mitsubishi PA-10 arm with 7 degrees of freedom (DOF) in non-stationary environments. Because of the specific kinematic model of the arm, path planning can be first reduced to a redundant 6-DOF problem in a 5D configuration space, which can be further decomposed into two problems: (i) 3D position planning in Cartesian space and (ii) planning in a 3D space composed of two orientation angles and an explicit parameterization of the arms redundancy. Position and orientation planning are interweaving and performed "on-the-fly" without explicit global knowledge of the environment using two instances of the dynamic wave expansion neural network (DWENN), an effective method for path generation in arbitrarily changing environments. The dynamic and explorative nature of the DWENN algorithm allows to treat stationary and dynamic obstacles in a unified manner. Through a number of simulative tests, we show that the planner is capable of reaching both a satisfactory robustness level and real-time performance, as required by many practical applications

We present a hybrid learning method bridging the fields of recurrent neural networks, unsupervised Hebbian learning, vector quantization, and supervised learning to implement a sophisticated image and feature segmentation architecture. This architecture is based on the competitive layer model (CLM), a dynamic feature binding model, which is applicable on a wide range of perceptual grouping and segmentation problems. A predefined target segmentation can be achieved as attractor states of this linear threshold recurrent network, if the lateral weights are chosen by Hebbian learning. The weight matrix is given by the correlation matrix of special pattern vectors with a structure dependent on the target labeling. Generalization is achieved by applying vector quantization on pair-wise feature relations, like proximity and similarity, defined by external knowledge. We show the successful application of the method to a number of artifical test examples and a medical image segmentation problem of fluorescence microscope cell images.

We present an on-line decentralized approach to collision-free path planning for two robot arms. During the real-time planning, each arm represents a dynamic obstacle for another one, which allows to treat the motion of the latter, as well as the motion of other objects in the workspace in a unified fashion. The motion for each arm is planned independently, and the only information which is "shared" is the intended configuration of each robot. The approach relies therefore exclusively on the dynamic, explorative path generation, which is performed using the dynamic wave expansion neural network. Our simulative experiments for the case of two robot arms with 3-DOFs in 3D reveal that the proposed approach, without any complicated heuristics, any priority assignment, and any global optimization of an objective cost function, is capable of producing feasible paths "on-the-fly". The robustness and efficiency of the method are demonstrated statistically through a number of random tests.

We introduce a new learning rule for fully recurrent neural networks which we call backpropagation-decorrelation rule (BPDC). It combines important principles: one-step backpropagation of errors and the usage of temporal memory in the network dynamics by means of decorrelation of activations. The BPDC rule is derived and theoretically justified from regarding learning as a constraint optimization problem and applies uniformly in discrete and continuous time. It is very easy to implement, and has a minimal complexity of 2N multiplications per time-step in the single output case. Nevertheless we obtain fast tracking and excellent performance in some benchmark problems including the Mackey-Glass time-series.

In this paper, we present a neural network model that realizes a dynamic version of the distance transform algorithm (used for path planning in a stationary domain). The novel version is capable of performing path generation for highly dynamic environments. The neural network has discrete-time dynamics, is locally connected, and, hence, computationally efficient. No preliminary information about the world status is required for the planning process. Path generation is performed via the neural-activity landscape, which forms a dynamically-updating potential field over a distributed representation of the configuration space of a robot. The network dynamics guarantees local adaptations and includes a set of strict rules for determining the next step in the path for a robot. According to these rules, planned paths tend to be optimal in a L/sub 1/ metric. Simulation results in a series of experiments for various dynamical situations prove the effectiveness of the proposed model.

This paper presents two contributions: (i) a new type of neural network the dynamic wave expansion neural network, for path generation in a dynamic environment for both mobile robots and robotic manipulators, and (ii) the simulative comparisons to known discrete-time neural network models - the classical resistive grid model, and the Hopfield-type neural network, proposed by Glasius et al. The network has discrete-time dynamics, is locally connected, highly parallel, and hence, computationally efficient. The model does not require any a-priory information about the environment. The path is generated according to a neural-activity landscape, which forms a dynamically updating scalar potential field over a distributed representation of the configuration space of a robot. The simulations reveal that the proposed model yields dominantly shorter paths, especially in highly-dynamic environments.

A major challenge for the realization of intelligent robots is to supply them with cognitive abilities in order to allow ordinary users to program them easily and intuitively. One approach to such programming is teaching work tasks by interactive demonstration. To make this effective and convenient for the user, the machine must be capable of establishing a common focus of attention and be able to use and integrate spoken instructions, visual perception, and non-verbal clues like gestural commands. We report progress in building a hybrid architecture that combines statistical methods, neural networks, and finite state machines into an integrated system for instructing grasping tasks by man-machine interaction. The system combines the GRAVIS-robot for visual attention and gestural instruction with an intelligent interface for speech recognition and linguistic interpretation, and a modality fusion module to allow multi-modal task-oriented man-machine communication with respect to dextrous robot manipulation of objects.

A major goal for the realization of a new generation of intelligent robots is the capability of instructing work tasks by interactive demonstration. To make such a process efficient and convenient for the human user requires that both the robot and the user can establish and maintain a common focus of attention. We describe a hybrid architecture that combines neural networks and finite stale machines into a flexible framework for controlling the behaviour of a vision based robot called GRAVIS-robot (Gestural Recognition Active Vision System robot). It consists of a binocular camera head, a 6 DOF robot arm and a 9 DOF multifingered hand. We focus primarily on nonverbal communication based on gestural commands of a human instructor which will at a later stage be complemented by spoken instructions.

We present a recurrent neural network for feature binding and sensory segmentation: the competitive-layer model (CLM). The CLM uses topographically structured competitive and cooperative interactions in a layered network to partition a set of input features into salient groups. The dynamics is formulated within a standard additive recurrent network with linear threshold neurons. Contextual relations among features are coded by pairwise compatibilities, which define an energy function to be minimized by the neural dynamics. Due to the usage of dynamical winner-take-all circuits, the model gains more flexible response properties than spin models of segmentation by exploiting amplitude information in the grouping process. We prove analytic results on the convergence and stable attractors of the CLM, which generalize earlier results on winner-take-all networks, and incorporate deterministic annealing for robustness against local minima. The piecewise linear dynamics of the CLM allows a linear eigensubspace analysis, which we use to analyze the dynamics of binding in conjunction with annealing. For the example of contour detection, we show how the CLM can integrate figure-ground segmentation and grouping into a unified model.

We show that using recurrent networks with finite time constants is not contradictory to arbitrary fast signal transmission in a closed loop with appropriate feedbacks. This surprising result is due to the occurrence of infinitely amplifying subloops, which we formally describe by differential inclusions. The theory then shows that the transmission speed depends crucially on the gain of the transfer function. Generalising the theoretical framework we demonstrate how to build efficient, fast and robust neuro-controllers with pre-specified performance by application to the benchmark problem of balancing the inverted pendulum.