D&A turns any VLM into a zero-shot robot reward model with SOTA on three benchmarks
KyleMorgenstein · x · 2026-11-04
- D&A (Decoupling Description & Assessment) improves any VLM family — GPT, Gemini, Claude, Qwen — as a zero-shot reward model for robotics, scaling with base model gains.
- It enables zero-shot RL and filtered IL, improves code-as-policies via rich language feedback, and hits SOTA on three benchmarks.
- Cited example: densifying a long-horizon sparse-reward NIST assembly task using per-task reward bumps from rollout video — "dynamics reward models, here we come."
- AlphaNIST: AlphaZero-style end-to-end policies learn contact-rich precision assembly in sim from a binary success reward alone — 20 tasks, parts as small as 4mm, no prescribed manipulation stages, training across heterogeneous assets and cached partial-assembly starts.
Related event: D&A Turns Any VLM into a Zero-Shot Robot Reward Model(3 posts)→
More from Embodied
- Open-weight LLMs control robots zero-shot: Qwen scores 31.39 on RoboDojo, beating GPT-6 Astra — pabbeel · 2026-11-04
- Apple reportedly beta tests Apple Watch AI feature that analyzes nearby conversations — Polymarket · 2026-11-04
- Boston Dynamics rebuilt Stretch Gen 2 from 100,000 operating hours, cutting 81% of internal cabling — xmercury_one · 2026-11-04
- Robotics funding hits another all-time high as capital keeps pouring in — Fowe · 2026-11-04
- Robot learns mobility and manipulation with just 15 minutes of training data — GordonWetzstein · 2026-11-04
- Zero-shot mobile manipulation demo stacks Dream-1 policy, MCP planner and ROS nav — scott_e_reed · 2026-11-04