Uncertainty-Aware Robotic World Model Makes Offline MBRL Work on Real Robots

Authors: Chenhao Li, Andreas Krause, Marco Hutter

Institution: ETH Zurich

arXiv: 2504.16680

Abstract

RWM-U extends autoregressive robotic world models with epistemic uncertainty estimation via ensembles, enabling temporally consistent multi-step rollouts with uncertainty propagated over long horizons. Combined with MOPO-PPO (uncertainty-penalized policy optimization), it enables training effective policies entirely from offline datasets on real quadruped and humanoid robots.

Key Method

  • RWM-U: autoregressive world model with ensemble-based epistemic uncertainty
  • Uncertainty signal penalizes high-risk imagined transitions during policy learning
  • MOPO-PPO: adapts offline model-based RL with PPO for stable on-policy optimization
  • Handles compounding errors and distribution shift in long-horizon rollouts

Safety Relevance

Uncertainty penalization is the core safety mechanism: by making the policy pessimistic about regions where the world model is unreliable, the agent avoids extrapolating into unsafe states not covered by offline data.

Results

  • Evaluated on manipulation and locomotion tasks (simulation + real hardware)
  • Outperforms model-free and uncertainty-unaware baselines
  • Fusing real-world data with simulation further improves robustness

Code

GitHub: leggedrobotics/robotic_world_model