SkeLAM grounds robot world models in skeleton motion
Qi Lv and five coauthors presented SkeLAM, a skeleton-grounded latent-action model for controllable robot world modeling. It separates robot kinematics from residual visual dynamics and uses separate conditioning pathways to generate future observations from the current scene and robot actions. The paper reports that, on DROID, SkeLAM reduced end-effector ADE by 19.5% and LPIPS by 11.1% versus DreamDojo, while qualitative rollouts demonstrated zero-shot human-to-robot motion transfer.
SkeLAM addresses a central challenge in video-trained robot world models by improving the correspondence between learned latent actions and robot controls, although its reported results come from the authors’ paper.
Computer Science > Robotics
arXiv:2610.31474 (cs)
[Submitted on 22 Oct 2026]
Title:Skeleton-Grounded Latent Action Model for Robot World Models
Authors:Qi Lv, Jiayi Lin, Yuxin Jiang, Jianzong Wang, Xiang Deng, Mike Shou Zheng
View a PDF of the paper titled Skeleton-Grounded Latent Action Model for Robot World Models, by Qi Lv and 5 other authors