Open-weight LLMs control robots zero-shot: Qwen scores 31.39 on RoboDojo, beating GPT-6 Astra
pabbeel · x · 2026-11-04
- The OpenRUA team released uni-robot-agent, a unified modular harness that lets general-purpose LLMs act as robot policies without privileged inputs, spanning simulation (RoboDojo, LIBERO) and physical arms via shared perception, motion, skills, and episode-control tools.
- Key finding: robot-use is a general capability that emerges across many LLMs—including open-weight models like Qwen and DeepSeek—without RSI, distillation, fine-tuning, or robot-specific data. Models approach tasks like math problems: understand the scene, triangulate feature points across views, then compute grasp poses via coordinate transforms.
- Headline numbers: Qwen3.8-Flash-Next scores 31.39 zero-shot on RoboDojo (35.89 with depth), beating GPT-6 Astra's official 28.97; with the harness Astra reaches 80.50. On 19 real-world tasks Qwen scores 53.50/100 zero-shot, outperforming fine-tuned π₀.₅ (47.75) and Dream-1 (43.50) at 1/19.5× the cost of Astra.
- Failure-mode analysis shows RoboDojo scores correlate strongly with general visual understanding, spatial grounding, geometric coding, embodied reasoning, and benchmarks like ERQA and MMMU-Pro, suggesting robot-use capability emerges naturally with general capability.
More from Embodied
- Berkeley researchers use RL to improve robot data collection with Data Inoculation — berkeley_ai · 2026-11-04
- Shield AI unveils X-BAT 001, full-scale tail-sitting VTOL stealth combat drone prototype — BrettKrieger12 · 2026-11-04
- OpenCV Live demos robot visual cortex with RealSense, Jetson, ROS 2 and VLMs — chrismatthieu · 2026-11-04
- Flipper One opens pre-orders ahead of Kickstarter campaign — user_7832 · 2026-11-04
- Sutton shares RL-trained robot dog preparing for lunar terrain — RichardSSutton · 2026-11-04
- Apple reportedly beta tests Apple Watch AI feature that analyzes nearby conversations — Polymarket · 2026-11-04