Event

FOCUS: Optimal Transport-Supervised Observation-Only Imitation Learning

1 reports · 1 sources

João A. Cândido Ramos and other authors propose the FOCUS imitation learning framework, which combines local optimal transport alignment with short-horizon differentiable actor-critic learning for observation-only imitation learning. The authors state that the method does not require time-aligned expert trajectories, nor expert-aligned initialization or teacher forcing; it uses a soft best-of-K objective to match agent rollout windows to multiple sampled expert segments, with the transport cost providing a differentiable imitation reward, and uses a learned critic to bootstrap values beyond the rollout horizon. In contact-rich locomotion tasks in DFlex (including a musculoskeletal humanoid controlled by 152 muscle activations), the authors report that FOCUS outperforms adversarial observation-only imitation methods and prior differentiable imitation baselines in sample efficiency and training stability.