D&A Turns Any VLM into a Zero-Shot Robot Reward Model, SOTA on 3 Benchmarks
DJiafei · x · 2026-11-03
UW and NVIDIA researchers introduce D&A (Decouple Description and Assessment), which splits VLM reward estimation into describing what happened, then assessing progress against explicit success criteria — fixing confirmation bias and hallucinated progress in zero-shot methods like GVL/TOPReward. Training-free (two prompts + an API), it scales across GPT, Gemini, Claude and Qwen (28 models), enables zero-shot RL and filtered IL, improves code-as-policies with rich language feedback, and hits SOTA on three reward benchmarks. Code and benchmark data are open.
More from Embodied
- Robot "physics problem" stems from training data, not the model, says researcher — chris_j_paxton · 2026-11-03
- Neuracore ships Trace Viewer: full visibility into robot data collection pipelines — stepjamUK · 2026-11-03
- UK robotics brain drain is reversing: ROSCon UK booth flooded with job seekers — stepjamUK · 2026-11-03
- Nature: Connectome simulations identify a three-neuron circuit generating fly walking rhythms — patrickmineault · 2026-11-03
- Duke's Humanoid V2 robot has two RealSense-powered heads and no front or back — chrismatthieu · 2026-11-03
- Bullish take: Physical AI will be 1,000,000x bigger than Digital AI — alexmacgregor__ · 2026-11-03