D&A turns any VLM into a zero-shot robot reward model, hitting SOTA on 3 benchmarks
KyleMorgenstein · x · 2026-11-04
Researchers introduce D&A (Decoupling description & assessment), which improves GPT, Gemini, Claude and Qwen VLMs as zero-shot robot reward models with no training. Feeding a video of any robot task plus a text description yields an off-the-shelf reward model, enabling zero-shot RL and filtered IL, improving code-as-policies via rich language feedback, and achieving SOTA on three benchmarks. Demo shows D&A only grants extra reward when the goal is visible in the image.
Related event: D&A Turns Any VLM into a Zero-Training Robot Reward Model(4 posts)→
More from Embodied
- Corebrum OS: open-source Physical AI OS now installable on NVIDIA Jetson and any Ubuntu machine — chrismatthieu · 2026-11-04
- Artificial Analysis launches AA-Robotics benchmark: GPT-6 Astra tops robot arm control at $16.36 per task — ArtificialAnlys · 2026-11-04
- Knowing when to stop is hard: Claude Opus 5.5 pushed a crocodile for 22 extra minutes — ArtificialAnlys · 2026-11-04
- Waymo expands South Bay coverage with Berryessa and Milpitas BART connections — reed · 2026-11-04
- Diamandis: a household robot saving 10 hours a week — 520 hours a year — is the robot economy people care about — PeterDiamandis · 2026-11-04
- Wafer Systems founder joins Nothing to build AI-native mobile OS after two years on the vision — zack_overflow · 2026-11-04