Preprint
Jul 2026
Local Motion Matters: A Deconstruct-Recompose Paradigm for Reinforcement Learning Pre-training from Videos
This work proposes a novel Deconstruct-Recompose Paradigm (DRP) for learning transferable local motion representations and introduces a Dual-Attention Encoder to learn local motion representations from these Atomic Actions, capturing their spatiotemporal relationships.
Jinwen Wang, Youfang Lin, X. Hu et al.
· 0 citations