Intelligent robots cannot be programmed in advance for all possible situations, but they should be able to generalize based on the acquired knowledge. In robot learning based on imitation of human activity we often use statistical methods that generalize observed (learned) movements. The acquired data is used to generate useful robots responses in situations for which the robot has not been specifically instructed how to respond. The paper describes the robot learning with Gaussian process regression that creates the model and estimates the parameters for generalization of the acquired motor knowledge, which is accumulated as a database of example movements. New actions are synthesized by applying Gaussian process regression, where the goal and other characteristics of an action are utilized as queries to create an optimal control policy with respect to the previously acquired knowledge. The paper demonstrates that the proposed methodology can be integrated with an active vision system of a humanoid robot. 3D vision data is used to provide query points for statistical generalization.

The goal of this paper is to investigate how to acquire useful action knowledge by observing the results of exploratory actions on objects. We focus on poking as a representative type of nonprehensile manipulation. Poking can be defined as a short term pushing action. Here we propose an explorative process that allows the robot to learn the relationship between the point of contact on the object boundary and the angle of poke and the actual response of an object. The robot acquires this knowledge without having any prior knowledge about the action. Initially, the robot was only able to move in random directions. Such self emergent processes are essential for the early cognition.The proposed process has been implemented and tested on the humanoid robot Hoap-3.

We present a new learning framework for synthesizing goal-directed actions from example movements. The approach is based on the memorization of training data and locally weighted regression to compute suitable movements for a large range of situations. The proposed method avoids making specific assumptions about an adequate representation of the task. Instead, we use a general representation based on fifth order splines. The data used for learning comes either from the observation of events in the Cartesian space or from the actual movement execution on the robot. Thus it informs us about the appropriate motion in the example situations. We show that by applying locally weighted regression to such data, we can generate actions having proper dynamics to solve the given task. To test the validity of the approach, we present simulation results under various conditions as well as experiments on a real robot.