ELA: End-to-End Latent Action Policy Learning from Video Demonstrations

This research introduces ELA, a framework for learning robot policies from video demonstrations by jointly optimizing latent actions for both policy learning and action decoding. The method addresses limitations in existing latent-action approaches that fail to optimize for downstream control. Experiments demonstrate consistent performance improvements over prior baselines on both simulated and real-world manipulation tasks. The framework integrates with existing imitation learning methods using lightweight 4M-parameter decoders without requiring changes to policy architectures or training objectives.