SkeLAM Uses Skeleton-Grounded Latent Actions to Decouple Motion from Appearance in Robot World Models
jiqizhixin · x · 2026-11-11
SkeLAM, from HIT Shenzhen, NUS Show Lab and Ping An Technology, tackles the entanglement of intended movement with appearance changes in video world models.
- Bottleneck: latent actions learned directly from video entangle motion with texture, background and viewpoint, making cross-embodiment transfer hard
- Mechanism: skeleton supervision anchors a kinematic latent component while a separate residual captures visual dynamics; cross-modal swapping/reconstruction encourages alignment; an adapter maps robot controls into kinematic space with the residual zeroed for action-conditioned inference
- Results: fine-tunes a pretrained Cosmos-Predict2.5-2B backbone with roughly 10 hours of robot data; lower average displacement error on DROID evaluation
More from Embodied
- Shepherd Robotics details Shepherd-0, its physical skill learning system built for robot deployment — KlausCodes · 2026-11-11
- MacroHands debuts: 3D hand tracking from egocentric video, stereo variant cuts depth error to 13.1mm — eliebakouch · 2026-11-11
- US robotaxi deployments are closing the gap with China, new data shows — skorusARK · 2026-11-11
- OpenTac debuts at CoRL as researchers call tactile the missing half of robot data — anand_bhattad · 2026-11-11
- Fully DIY in-wheel-motor autonomous RC car runs SLAM on a ¥2000 LiDAR — ___Mufasaa · 2026-11-11
- AHA 2026: AI reads smartwatch ECGs for heart failure and speeds echo interpretation — AmiBhattMD · 2026-11-11