New RL Pre-training Method Improves Transferability with Local Motion.

Pre-training reinforcement learning (RL) models using large video datasets holds significant potential for boosting efficiency, but current methods often struggle with transferability across different domains. This is because they typically model an agent's motion globally, which becomes tightly coupled with its specific physical form. Researchers propose a new approach called the Deconstruct-Recompose Paradigm (DRP) to overcome this limitation by focusing on local motion patterns. DRP operates in two phases. First, the "Deconstruct" phase identifies and tracks multiple local points on an agent, treating their frame-wise movements as "Atomic Actions." A Dual-Attention Encoder then learns representations of these local motions, capturing their spatial and temporal relationships. In the "Recompose" phase, these local motion representations are combined with a learnable Motion Aggregation Token via a latent dynamics model, and an adapter helps bridge these local motions to specific downstream actions, accelerating policy learning. Experiments show DRP significantly improves sample efficiency and performance in various robotic control and manipulation tasks.