Event

SkeLAM: Skeleton-Guided Robotic World Model Proposed

1 reports · 1 sources

Researchers propose SkeLAM, a skeleton-guided latent action model for controllable robot world modeling. The method decomposes latent actions into a kinematic component supervised by the robot skeleton and a residual component that preserves complementary visual dynamics, and enables interchangeability between RGB and skeleton kinematic representations through cross-modal swap reconstruction. On the DROID benchmark, SkeLAM reduces end-effector ADE by 19.5% and LPIPS by 11.1% compared with DreamDojo. Policy evaluation shows that its predicted success rates are consistent with results measured from real robot execution.