Learning Robust Visuomotor Policies via Dynamics Alignment

This research proposes DAP, a policy framework that couples action generation with an action-conditioned dynamics model to improve robustness in visuomotor tasks. By co-sampling actions and future observations through flow extrapolation, DAP achieves the highest average success rate across four simulated and three real-world tasks, with relative improvements of 56.5% in simulation and 33.3% in the real world over the strongest baseline. The method also recovers self-consistent action-future pairs when future-observation estimates are corrupted.