Storm: An Experience-Based Framework for Robot Learning From Demonstration

Abstract

Learning from demonstration (LfD) can be used to increase the behavioural repertoire of a robot, but most demonstration-based learning techniques do not enable a robot to acquire knowledge about the limitations of its own body and use that information during learning. In this letter, we propose Storm, an LfD framework that enables acquiring trajectories in high-dimensional spaces, incorporates collision awareness, and can be adapted to different robots. Storm combines a collection of modules: i) robot embodiment exploration using motor babbling in order to acquire knowledge about the robot's own body, stored in the form of joint-specific graphs that encode reachable points and reachability constraints, ii) human-robot model mapping based on which human skeleton observations are mapped to the robot's embodiment, and iii) demonstration-based trajectory learning and subsequent reproduction of the learned actions using Gaussian mixture regression. We validate various aspects of our approach experimentally: i) exploration with different numbers of babbling points for three distinct robots, ii) path planning performance, including in the presence of obstacles, and iii) the acceptance of reproduced trajectories through a small-scale, real-world user study. The results demonstrate that Storm can produce versatile behaviours on different robots, and that trajectory reproductions are generally rated well by external observers, which is important for overall user acceptance.