Learning Generalizable Visuomotor Policy through Dynamics-Alignment
This work addresses generalization issues in robot manipulation for robotics, but it is incremental as it builds on existing video prediction and behavior cloning methods.
The paper tackles the problem of poor generalization in robot learning from demonstrations by proposing a Dynamics-Aligned Flow Matching Policy (DAP) that integrates dynamics prediction into policy learning, resulting in superior generalization performance on real-world robotic manipulation tasks, particularly in out-of-distribution scenarios like visual distractions and lighting variations.
Behavior cloning methods for robot learning suffer from poor generalization due to limited data support beyond expert demonstrations. Recent approaches leveraging video prediction models have shown promising results by learning rich spatiotemporal representations from large-scale datasets. However, these models learn action-agnostic dynamics that cannot distinguish between different control inputs, limiting their utility for precise manipulation tasks and requiring large pretraining datasets. We propose a Dynamics-Aligned Flow Matching Policy (DAP) that integrates dynamics prediction into policy learning. Our method introduces a novel architecture where policy and dynamics models provide mutual corrective feedback during action generation, enabling self-correction and improved generalization. Empirical validation demonstrates generalization performance superior to baseline methods on real-world robotic manipulation tasks, showing particular robustness in OOD scenarios including visual distractions and lighting variations.