Learning Domain Randomization Distributions for Training Robust Locomotion Policies

摘要

This paper considers the problem of learning behaviors in simulation without knowledge of the precise dynamical properties of the target robot platform(s). In this context, our learning goal is to mutually maximize task efficacy on each environment considered and generalization across the widest possible range of environmental conditions. The physical parameters of the simulator are modified by a component of our technique that learns the Domain Randomization (DR) that is appropriate at each learning epoch to maximally challenge the current behavior policy, without being overly challenging, which can hinder learning progress. This so-called sweet spot distribution is a selection of simulated domains with the following properties: 1) The trained policy should be successful in environments sampled from the domain randomization distribution; and 2) The DR distribution made as wide as possible, to increase variability in the environments. These properties aim to ensure the trajectories encountered in the target system are close to those observed during training, as existing methods in machine learning are better suited for interpolation than extrapolation. We show how adapting the DR distribution while training context-conditioned policies results in improvements on jump-start and asymptotic performance when transferring a learned policy to the target environment <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> .