Jochen J. Steil

Backpropagation-decorrelation: online recurrent learning with O(N) complexity

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.

Improving reservoirs using intrinsic plasticity

The benefits of using intrinsic plasticity (IP), an unsupervised, local, biologically inspired adaptation rule that tunes the probability density of a neurons output towards an exponential distribution-thereby realizing an information maximization-have already been demonstrated. In this work, we extend the ideas of this adaptation method to a more commonly used non-linearity and a Gaussian output distribution. After deriving the learning rules, we show the effects of the bounded output of the transfer function on the moments of the actual output distribution. This allows us to show that the rule converges to the expected distributions, even in random recurrent networks. The IP rule is evaluated in a reservoir computing setting, which is a temporal processing technique which uses random, untrained recurrent networks as excitable media, where the networks state is fed to a linear regressor used to calculate the desired output. We present an experimental comparison of the different IP rules on three benchmark tasks with different characteristics. Furthermore, we show that this unsupervised reservoir adaptation is able to adapt networks with very constrained topologies, such as a 1D lattice which generally shows quite unsuitable dynamic behavior, to a reservoir that can be used to solve complex tasks. We clearly demonstrate that IP is able to make reservoir computing more robust: the internal dynamics can autonomously tune themselves-irrespective of initial weights or input scaling-to the dynamic regime which is optimal for a given task.

Goal Babbling Permits Direct Learning of Inverse Kinematics

We present an approach to learn inverse kinematics of redundant systems without prior- or expert-knowledge. The method allows for an iterative bootstrapping and refinement of the inverse kinematics estimate. The essential novelty lies in a path-based sampling approach: we generate training data along paths, which result from execution of the currently learned estimate along a desired path towards a goal. The information structure thereby induced enables an efficient detection and resolution of inconsistent samples solely from directly observable data. We derive and illustrate the exploration and learning process with a low-dimensional kinematic example that provides direct insight into the bootstrapping process. We further show that the method scales for high dimensional problems, such as the Honda humanoid robot or hyperredundant planar arms with up to 50 degrees of freedom.

A Competitive-Layer Model for Feature Binding and Sensory Segmentation

We present a recurrent neural network for feature binding and sensory segmentation: the competitive-layer model (CLM). The CLM uses topo-graphically 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.

Where to Look Next? Combining Static and Dynamic Proto-objects in a TVA-based Model of Visual Attention

To decide “Where to look next ?” is a central function of the attention system of humans, animals and robots. Control of attention depends on three factors, that is, low-level static and dynamic visual features of the environment (bottom-up), medium-level visual features of proto-objects and the task (top-down). We present a novel integrated computational model that includes all these factors in a coherent architecture based on findings and constraints from the primate visual system. The model combines spatially inhomogeneous processing of static features, spatio-temporal motion features and task-dependent priority control in the form of the first computational implementation of saliency computation as specified by the “Theory of Visual Attention” (TVA, [7]). Importantly, static and dynamic processing streams are fused at the level of visual proto-objects, that is, ellipsoidal visual units that have the additional medium-level features of position, size, shape and orientation of the principal axis. Proto-objects serve as input to the TVA process that combines top-down and bottom-up information for computing attentional priorities so that relatively complex search tasks can be implemented. To this end, separately computed static and dynamic proto-objects are filtered and subsequently merged into one combined map of proto-objects. For each proto-object, attentional priorities in the form of attentional weights are computed according to TVA. The target of the next saccade is the center of gravity of the proto-object with the highest weight according to the task. We illustrate the approach by applying it to several real world image sequences and show that it is robust to parameter variations.

Platform portable anthropomorphic grasping with the bielefeld 20-DOF shadow and 9-DOF TUM hand

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.

Multi-modal human-machine communication for instructing robot grasping tasks

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.

Task-oriented quality measures for dextrous grasping

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

Situated robot learning for multi-modal instruction and imitation of grasping

Abstract A key prerequisite to make user instruction of work tasks by interactive demonstration effective and convenient is situated multi-modal interaction aiming at an enhancement of robot learning beyond simple low-level skill acquisition. We report the status of the Bielefeld GRAVIS-robot system that combines visual attention and gestural instruction with an intelligent interface for speech recognition and linguistic interpretation to allow multi-modal task-oriented instructions. With respect to this platform, we discuss the essential role of learning for robust functioning of the robot and sketch the concept of an integrated architecture for situated learning on the system level. It has the long-term goal to demonstrate speech-supported imitation learning of robot actions. We describe the current state of its realization to enable imitation of human hand postures for flexible grasping and give quantitative results for grasping a broad range of everyday objects.

Efficient Exploratory Learning of Inverse Kinematics on a Bionic Elephant Trunk

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.