Stabilized Likelihood-based Imitation Learning via Denoising Continuous Normalizing Flow
State-of-the-art imitation learning (IL) approaches, e.g, GAIL, apply adversarial training to minimize the discrepancy between expert and learner behaviors, which is prone to unstable training and mode collapse. In this work, we propose SLIL – Stabilized Likelihood-based Imitation Learning – a novel IL approach that directly maximizes the likelihood of observing the expert demonstrations. SLIL is a two-stage optimization framework, where in stage one the expert state distribution is estimated via a new method for denoising continuous normalizing flow, and in stage two the learner policy is trained to match both the expert’s policy and state distribution. Experimental evaluation of SLIL compared with several baselines in ten different physics-based control tasks reveals superior results in terms of learner policy performance, training stability, and mode distribution preservation.
PDF Abstract
