Σ-LeWM
Learning world models via S-information for robot control (2026 preprint)
Σ-LeWM: Learning World Models via S-Information for Robot Control — Xinrui Zu, Michael Przystupa, Zhao Yang, Shujian Yu, and Kevin Sebastian Luck. Preprint, 2026.
Project website · Code · Hugging Face
Σ-LeWM learns a latent world model through three views of a transition: predicting the next state, explaining the current state from the next state and action, and recovering the action between consecutive states. Together, forward, backward, and inverse dynamics encourage representations that preserve the structure needed for robot control.
The two auxiliary predictors are used during training. At test time, the model retains LeWM’s forward dynamics and CEM planner. We evaluate control performance and planning efficiency on Push-T, TwoRoom, Reacher, OGBench-Cube, and OGBench-Scene, alongside Gear-Assembly and Cube-Stacking in IsaacSim.
The arXiv-version main table reports 94.5% success on Push-T and 80.7% on OGBench-Scene, compared with 90.0% and 57.7% for LeWM. These are means over 3 training seeds × 5 evaluation seeds × 50 episodes; the full table includes standard deviations and all baselines.