ETH's Uncertainty-Aware World Model Makes Offline RL Work on Real Robots
breadli428 · x · 2026-10-15
ETH Zurich (Chenhao Li, Andreas Krause, Marco Hutter) released RWM-U, an uncertainty-aware robotic world model that makes offline model-based RL practical on physical robots.
- Problem: offline MBRL is data-efficient but suffers from compounding errors and distribution shift in long-horizon rollouts, and mostly worked only in simulation benchmarks rather than on noisy, biased, partially observed real-robot data.
- Method: RWM-U augments autoregressive world models with epistemic uncertainty estimation for temporally consistent multi-step rollouts, combined with MOPO-PPO, adapting uncertainty-penalized policy optimization to the stable on-policy PPO framework.
- Results: Across manipulation and locomotion tasks in sim and on a real quadruped and humanoid, policies trained purely from offline datasets outperform model-free and uncertainty-unaware model-based baselines; fusing real-world data even surpasses online model-free baselines trained in sim.
Paper, code and poster are available.
More from Embodied
- Rhoda unveils fingertip-like tactile sensor letting robots identify objects while blindfolded — CSProfKGD · 2026-10-15
- Infinite Machine's Olto e-scooter spotted roaming NYC streets twice in 4 hours — soleio · 2026-10-15
- Vals AI Puts Frontier Models in Real-Time MuJoCo Robot Simulations — scaling01 · 2026-10-15
- Drone Footage Reveals Tesla's Optimus Factory at Giga Texas, 40% Steel Assembled — Scobleizer · 2026-10-15
- When will a humanoid robot team beat the World Cup champs? Reliability still at 80% on easy tasks — mathemagic1an · 2026-10-15
- UPenn lab shows GPT-6 Astra can pilot FPV drones from raw images, no mapping stack needed — antoniloq · 2026-10-15