arXiv cs.AIPaper
EgoLAP: Learning from Egocentric Human Data through Language-Action Reasoning
The real win is that human video becomes useful robot training data without costly robot demonstrations. By operating at the level of motion intent rather than low-level actions, EgoLAP sidesteps the embodiment mismatch that made naïve imitation learning fail. For builders working on embodied agents, this suggests that language-abstracted representations are the right abstraction layer for cross-embodiment transfer. The 80% real-world task success rate signals this is moving beyond simulation validation.