Model-Free Apprenticeship Learning for Transfer of Human Impedance Behaviour

We present a method for transferring behaviour from humans to robots via apprenticeship learning. While previous methods have relied on an accurate model of the demonstrator's dynamics, in most practical settings such models fail to capture (i) complex, non-linear dynamics of the human musculoskeletal system, and (ii) inconsistencies between modelling assumptions and the configuration and placement of measurement apparatus. To avoid such issues, we propose a model-free approach to apprenticeship learning, in which off- policy, model-free reinforcement learning techniques are used to extract a model of the objective function optimised in human behaviour. As a key ingredient, we derive a novel formulation of Least Squares Policy Iteration (LSPI) and Least Squares Temporal Difference learning (LSTD) to enable their application in this setting. The robustness of our approach is demonstrated in experiments where human hitting behaviour is transferred to a non-biomorphic robotic device.