OpenMOSS Open-Sources WCM: A World Critic Model for VLA Reinforcement Learning

机器之心 · wechat · 2026-08-13

The OpenMOSS team open-sourced WorldCriticModel (WCM) to address the limitation of Vision-Language-Action (VLA) models relying solely on single-frame observations for value estimation in RL. Since robot control is inherently a Partially Observable Markov Decision Process (POMDP), a single frame cannot reliably capture dynamic trends or contact states. WCM unifies world prediction and value learning by forcing the Critic to predict the next latent state while scoring.

Experiments show WCM achieves significant breakthroughs across multiple simulation benchmarks and 7 real-robot tasks. In ManiSkill, it boosted a VLA model's success rate from 0.78% to 98.7%. Furthermore, starting with only 1 demonstration, WCM surpassed the Full-SFT baseline (which uses 50 demos) after just 250 steps of RL fine-tuning. The model supports both On-policy and Off-policy pipelines, with code, data, and weights fully open-sourced.

Original post →

OpenMOSS Open-Sources WCM: A World Critic Model for VLA Reinforcement Learning

More from Embodied