D&A: Decoupling Description from Assessment Turns Any VLM into a Robot Reward Model
weichiuma · x · 2026-11-04
UW researchers found VLMs used as reward models suffer heavy confirmation bias, overestimating task success. Their fix, D&A, decouples describing what happened from assessing how good it is — like rubric-based rewards for LLMs. Works zero-shot across GPT, Gemini, Claude and Qwen families, enables zero-shot RL and filtered IL, boosts code-as-policies with language feedback, and hits SOTA on three benchmarks. A live demo accepts any video with just an OpenRouter key.
More from Embodied
- NVIDIA shows AI agents teaching a robot to walk an obstacle course in sim — tomaszbednarz · 2026-11-04
- NYC team behind 'an extension of your thinking' to start shipping Stream next week — pzakin · 2026-11-04
- Co-designed robot learning stack from scratch: full robot under $15,000 — ChongZzZhang · 2026-11-04
- Jitendra Malik proposes '5-year-old paired test' to deflate robot manipulation demos — BradPorter_ · 2026-11-04
- Dev revives a dusty Meta Portal with a coding agent; tip: let second-hand units 'decant' for days before login works — msg · 2026-11-04
- BBC Films Epia Neuro Brain-Computer Interface Exoskeleton Glove Restoring Hand Function — prokarurol · 2026-11-04