Latent Action as Intention Enables Efficient Future Imagination for World Action Models
LAWA keeps future-aware robot planning while cutting inference latency by 42.9%.
The paper proposes compact latent actions as a stand-in for generated future observations in world action models. In RoboCasa tests, LAWA reports 65.6% few-shot and 80.8% full-data average success, beating a matched Fast-WAM baseline by 9.6 and 4.5 points. It matches the authors’ Joint-WAM performance level while avoiding the future-video branch at inference. The authors also report competitive zero-shot robustness on LIBERO-Plus and stronger real-world task results. HF Daily Papers' note
The paper proposes compact latent actions as a stand-in for generated future observations in world action models. In RoboCasa tests, LAWA reports 65.6% few-shot and 80.8% full-data average success, beating a matched Fast-WAM baseline by 9.6 and 4.5 points. It matches the authors’ Joint-WAM performance level while avoiding the future-video branch at inference. The authors also report competitive zero-shot robustness on LIBERO-Plus and stronger real-world task results. HF Daily Papers' note
score 5