WCM boosts VLA robot RL with a world-model critic and 149-task wins
OpenMOSS-Team · hf · 2026-08-04
WCM adds a world-model-style critic to VLA robot RL
OpenMOSS proposes WCM (World Critic Model) for reinforcement learning post-training of Vision-Language-Action models. The core diagnosis is that existing critic methods usually rely on single-frame observations or single-frame backbone latents, which mismatches the partially observable nature of robot control.
What changes
- WCM uses a lightweight LeJEPA-style architecture.
- It jointly predicts future latent state and estimates value, so the critic learns temporal dynamics instead of only regressing scalar returns.
- The method plugs into both on-policy and off-policy pipelines.
- It is compatible with backbones such as Pi0, Pi0.5, and OpenVLA-OFT.
Results
- Experiments cover 149 tasks across four benchmarks.
- The paper reports consistent state-of-the-art performance in both in-distribution and out-of-distribution settings.
- The authors also validate the method on seven real-world manipulation tasks, showing stable deployment across varied settings.
More from Embodied
- Humanoid raises $152 million at a $1.35 billion valuation before beta deployments — VraserX · 2026-08-04
- A retina-inspired optical-flow sensor moves motion computation onto the chip — ssh4net · 2026-08-04
- Figure says it had to build its own humanoid supply chain from scratch — adcock_brett · 2026-08-04
- DARPA lift challenge pits a gas bicopter against a 32-prop electric design — GoddenThomas · 2026-08-04
- Blue River co-founder says robotics only matters when machines get used — davidyin44 · 2026-08-04
- Video generation can turn one SOP shoot into multilingual training clips and grid-inspection data — xiaohu · 2026-08-04