Event

SkeLAM: A Robot World Model with Skeleton-Supervised Latent Actions

1 reports · 1 sources

SkeLAM, a robot world model proposed by Harbin Institute of Technology (Shenzhen), Show Lab at the National University of Singapore, and other institutions, uses robot skeletons to supervise latent actions in videos, with the goal of achieving zero-shot cross-embodiment transfer. The model is built on the pretrained Cosmos-Predict2.5-2B and trained on a total of about 10 hours of robot interaction data; during transfer, all model parameters are frozen, with no additional fine-tuning for demonstrations. In the DROID evaluation, compared with DreamDojo, SkeLAM reduces end-effector trajectory error by 19.5% and the perceptual distance LPIPS by 11.1%. These results come from video generation evaluations. The report states that moving from generated results to real robots still requires testing more complex contacts and longer manipulation processes, and evaluating the actual help of the generated videos for policy learning and closed-loop execution.