Abstract — This work is concerned with a framework for visual object recognition in real world tasks. Our approach is motivated by biological findings of the representation of space around the body, the so-called peripersonal space. We show that the principles behind those findings can lead to a natural structuring of object recognition tasks in artificial systems. We demonstrate this by the supervised learning and recognition of 20 complexshaped objects from unsegmented visual input.

We present a neural network approach to early motor learning. The goal is to explore the needs for bootstrapping the control of hand movements in a biologically plausible learning scenario. The model is applied to the control of hand postures of the humanoid robot ASIMO by means of full upper body movements. For training, we use an efficient online scheme for recurrent reservoir networks consisting of supervised backpropagation-decorrelation output adaptation and an unsupervised intrinsic plasticity reservoir optimization. We demonstrate that the network can acquire accurate inverse models for the highly redundant ASIMO, applying bi-manual target motions and exploiting all upper body degrees of freedom. We show that very few, but highly symmetric training data is sufficient to generate excellent generalization capabilities to untrained target motions. We also succeed in reproducing real motion recorded from a human demonstrator, massively differing from the training data in range and dynamics. The demonstrated generalization capabilities provide a fundamental prerequisite for an autonomous and incremental motor learning in an developmentally plausible way. Our exploration process – though not yet fully autonomous – clearly shows that goal- directed exploration can, in contrast to “babbling” of joints angles, be done very efficiently even for many degrees of freedom and non-linear kinematic configurations as ASIMOs.

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 trainig data along pathes, 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.

Abstract—We present a method for learning potential-based policies from constrained motion data. In contrast to previous approaches to direct policy learning, our method can combine observations from a variety of contexts where different constraints are in force, to learn the underlying unconstrained policy in form of its potential function. This allows us to generalise and predict behaviour where novel constraints apply. As a key ingredient, we first create multiple simple local models of the potential, and align those using an efficient algorithm. We can then detect and discard unsuitable subsets of the data and learn a global model from a cleanly pre-processed training set. We demonstrate our approach on systems of varying complexity, including kinematic data from the ASIMO humanoid robot with 22 degrees of freedom. I.

Many everyday human skills can be framed in terms of performing some task subject to constraints imposed by the environment. Constraints are usually unobservable and frequently change between contexts. In this paper, we present a novel approach for learning (unconstrained) control policies from movement data, where observations come from movements under different constraints. As a key ingredient, we introduce a small but highly effective modification to the standard risk functional, allowing us to make a meaningful comparison between the estimated policy and constrained observations. We demonstrate our approach on systems of varying complexity, including kinematic data from the ASIMO humanoid robot with 27 degrees of freedom, and present results for learning from human demonstration.

We consider the problem of direct policy learning in situations where the policies are only observable through their projections into the null-space of a set of dynamic, non-linear task constraints. We tackle the issue of deriving consistent data for the learning of such policies and make two contributions towards its solution. Firstly, we derive the conditions required to exactly reconstruct null-space policies and suggest a learning strategy based on this derivation. Secondly, we consider the case that the null-space policy is conservative and show that such a policy can be learnt more easily and robustly by learning the underlying potential function and using this as our representation of the policy. 1

Abstract. Many everyday human skills can be framed in terms of performing some task subject to constraints imposed by the task or the environment. Constraints are usually not observable and frequently change between contexts. In this chapter, we explore the problem of learning control policies from data containing variable, dynamic and non-linear constraints on motion. We discuss how an effective approach for doing this is to learn the unconstrained policy in awaythatisconsistent with the constraints. Wethengoontodiscussseveral recent algorithms for extracting policies from movement data, where observations are recorded under variable, unknown constraints. We review a number of experiments testing the performance of these algorithms and demonstrating how the resultant policy models generalise over constraints allowing prediction of behaviour under unseen settings where new constraints apply. 1

Abstract — Many everyday human skills can be framed in terms of performing some task subject to constraints imposed by the environment. Constraints are usually unobservable and frequently change between contexts. In this paper, we present a novel approach for learning (unconstrained) control policies from movement data, where observations come from movements under different constraints. As a key ingredient, we introduce a small but highly effective modification to the standard risk functional, allowing us to make a meaningful comparison between the estimated policy and constrained observations. We demonstrate our approach on systems of varying complexity, including kinematic data from the ASIMO humanoid robot with 27 degrees of freedom. I.

Abstract — 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 bimanual 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. I.

Abstract — Humanoid robots are highly redundant systems with respect to the tasks they are asked to perform. This redundancy manifests itself in the number of degrees of freedom of the robot exceeding the dimensionality of the task. Traditionally this redundancy has been utilised through optimal control in the null-space. Some cost function is defined that encodes secondary movement goals and movements are optimised with respect to this function, subject to fulfilment of task constraints. Until now design of cost functions has been carried out on an ad-hoc basis and has required time-consuming hand-tuning to ensure that the desired (or acceptable) behaviour is realised. Here we present a novel approach for designing cost functions for optimal control in the null-space by exploiting recent advances in statistical machine learning. The behaviour of a (kinematically or dynamically controlled) mechanical system performing some task is observed and separated into task- and null-space components. The null-space component is then modelled as a first order differential equation with the cost as the independent variable. Numerical solution of this equation provides training data for a statistical learning algorithm that is used to build an open-form model of the cost function. Results are presented in which the reconstructed function is used to replace that of the original control scheme and the resultant behaviour, for the same set of tasks, is compared.