Abstract — 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. I.

One of the central themes in autonomous robot research concerns the question how visual images of body movements by others can be interpreted and related to one’s own body movements and to language describing these body movements. The discovery of mirror neurons has shown that there are brain circuits which become active both in the perception and the re-enactment of bodily gestures, although it is so far unclear how these circuits can form, i.e. how neurons become mirror neurons. We report here further progress with our robot experiments in which a group of autonomous robots play language games in order to coordinate their visual, motor and cognitive body image. We have shown that the right kind of semiotic dynamics can lead to the self-organisation of a successful communication system with which robots can ask each other to perform certain actions. The main contribution of this paper is to show that if the robot has the capacity to ‘imagine ’ the behavior of his own body through selfsimulation, he is better able to guess what action corresponds to a visual image produced by another robot and thus guess the meaning of an unknown word. This leads to a significant speed-up in the way individual agents are able to coordinate visual categories, motor behaviors and language.

We show that inverse kinematics of different tools can be efficiently learned with a single recurrent neural network. Our model exploits all upper body degrees of freedom of the Honda humanoid robot research platform. Both hands are controlled at the same time with parametrized tool geometry. We show that generalization both in space as well as across tools is possible from very few training data. The network even permits extrapolation beyond the training data. For training we use an efficient online scheme for recurrent reservoir networks utilizing supervised backpropagation-decorrelation (BPDC) output adaptation and an unsupervised intrinsic plasticity (IP) reservoir optimization.

To improve the flexibility of robotic learning, it is important to realize an ability to generate a hierarchical structure. This paper proposes a learning framework which can dynamically change the planning space depending on the structure of the task. Synchronous motion information is utilized to generate modes and different modes correspond to different hierarchical structure of the controller. This enables efficient task planning and control using lowdimensional space. An object manipulation task is tested as an application, where an object is found and used as a tool (or as a part of the body) to extend the ability of the robot. The proposed framework is expected to be a basic learning model to account for body image acquisition including tool affordances. 1.