Event
Paper on Latent World Action Model LaWAM Released
1 reports · 1 sources
Jialei Chen and other authors released a paper proposing the Latent World Action Model LaWAM: instead of reconstructing future videos, it exposes predicted dynamics to robot policies via compact latent visual subgoals. Its core LaWM is trained in the latent space of a pretrained visual foundation model and reuses that model's forward decoder to predict future observation features. The authors report that LaWAM achieves a 98.6% success rate on LIBERO and 91.22% on RoboTwin, and also achieves leading or comparable success rates on real robot manipulation tasks, while maintaining low-latency inference: each action chunk prediction takes 187 milliseconds, and the wall-clock latency is up to 24 times lower than pixel-space WAM. The paper has been accepted to CoRL 2026.
Report timeline
- LaWAM: Efficient Dynamics-Aware Robot Policies with Latent World Action Models
LaWAM proposes a Latent World Action Model that exposes predictive dynamics to robot policies through compact latent visual subgoals rather than reconstructed future video. Its core LaWM trains a latent action model in the latent space of a pretrained vision foundation model and reuses its forward decoder to predict future observation features.
arXiv cs.RO ·