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

We present an approach to learn inverse kinematics of redundant systems without prior- or expert-knowledge. The method allows for iterative bootstrapping and refinement the estimate. essential novelty lies in a path-based sampling approach: we generate training data along paths, which result from execution currently learned estimate desired path towards goal. information structure thereby induced enables efficient detection resolution inconsistent samples solely directly observable data....

We present an approach to learn the inverse kinematics of "bionic handling assistant"-an elephant trunk robot. This task comprises substantial challenges including high dimensionality, restrictive and unknown actuation ranges, nonstationary system behavior. use a recent exploration scheme, online goal babbling, which deals with these by bootstrapping adapting on fly. show success method in extensive real-world experiments robot, novel combination learning traditional feedback control....

We evaluate the use of continuum kinematics with constant curvature as a kinematic model for Festo's "Bionic Handling Assistant" (BHA). introduce new, elegant, and parameterless method to deal geometric singularities in stretched positions, which allows capture pure elongations that are not naturally expressed by toroidal deformations underlying assumption. The stability is shown numeric simulations. how well this describes BHA using real-world position measurements quantitative ground truth...

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

The recent advent of compliant and kinematically redundant robots poses new research challenges for human-robot interaction. While these provide a great degree flexibility the realization complex applications, gained generates need additional modeling steps definition criteria redundancy resolution constraining robot's movement generation. explicit such usually require experts to adapt generation subsystem. A typical way dealing with this configuration challenge is utilize kinesthetic...

Feed-forward model-based control relies on models of the controlled plant, e.g., in robotics accurate knowledge manipulator kinematics or dynamics. However, mechanical and analytical do not capture all aspects a plant's intrinsic properties there remain unmodeled dynamics due to varying parameters, friction soft materials. In this context, machine learning is an alternative suitable technique extract non-linear plant from data. fully data-based suffer inaccuracies as well are inefficient if...

We propose a new and efficient approach to compute task oriented quality measures for dextrous grasps. Tasks can be specified as single wrench applied, rough direction in form of cone, or complex polytope. Based on the linear matrix inequality formalism treat friction cone constraints we formulate respective convex optimization problems, whose solutions give maximal applicable together with needed contact forces. Numerical experiments show that application grasps many contacts is possible.

Recent advances in the field of humanoid robotics increase complexity tasks that such robots can perform. This makes it increasingly difficult and inconvenient to program these manually. Furthermore, robots, contrast industrial should distant future behave within a social environment. Therefore, must be possible extend robot's abilities an easy natural way. To address requirements, this work investigates topic imitation learning motor skills. The focus lies on providing robot with ability...

We present a model of online Goal Babbling for the bootstrapping sensorimotor coordination. By modeling infants' early goal-directed movements we show that inverse models can be bootstrapped within few hundred even in very high-dimensional spaces. Our thereby explains how infants might initially acquire reaching skills without need exhaustive exploration, and robots do so feasible way. learning closed loop with exploration allows substantial speed-ups and, systems, outperforms previously...

We present a recurrent neural network for feature binding and sensory segmentation: the competitive-layer model (CLM). The CLM uses topo-graphically structured competitive cooperative interactions in layered to partition set of input features into salient groups. dynamics is formulated within standard additive with linear threshold neurons. Contextual relations among are coded by pairwise compatibilities, which define an energy function be minimized dynamics. Due usage dynamical...

A major challenge for the realization of intelligent robots is to supply them with cognitive abilities in order allow ordinary users program easily and intuitively. One way such programming teaching work tasks by interactive demonstration. To make this effective convenient user, machine must be capable establish a common focus attention able use integrate spoken instructions, visual perceptions, non-verbal clues like gestural commands. We report progress building hybrid architecture that...

Robot learning by imitation requires the detection of a tutor's action demonstration and its relevant parts. Current approaches implicitly assume unidirectional transfer knowledge from tutor to learner. The presented work challenges this predominant assumption based on an extensive user study with autonomously interacting robot. We show that providing feedback, robot learner influences human movement demonstrations in process learning. argue robot's feedback strongly shapes how tutors signal...

We present Oncilla robot, a novel mobile, quadruped legged locomotion machine. This large-cat sized, 5.1 robot is one of kind recent, bioinspired class designed with the capability model-free control. Animal in rough terrain clearly shaped by sensor feedback systems. Results show that agile and versatile possible without sensory signals to some extend, tracking becomes robust when control added (Ajaoolleian 2015). By incorporating mechanical blueprints inspired from animals, observing...

Feed-forward control relies on accurate knowledge about the controlled plant, e.g. models of manipulator kinematics or dynamics. However, for many plants, mechanical do not capture all aspects a plant plant's intrinsic properties, soft materials, hardly allow exact and efficient modeling. In this context, machine learning is suitable technique to extract non-linear from data. The paper shows that feed-forward based inversion hybrid forward model comprising learned error can significantly...

This work presents a novel approach to handling epistemic uncertainty estimates with motivation from Bayesian linear regression. We propose treating the model-dependent variance in predictive distribution-commonly associated uncertainty-as model for underlying data distribution. Using high-dimensional random feature transformations, this allows computationally efficient, parameter-free representation of arbitrary distributions. assessing whether query point lies within distribution, which...