Tsinghua/Berkeley open-source uni-robot-agent: Qwen beats GPT-6 Astra at 6.6% the cost
青稞AI · wechat · 2026-11-03
A Tsinghua + UC Berkeley team (corresponding authors include Ding Ning, Pieter Abbeel, Liu Zhiyuan, Xu Mengdi) released uni-robot-agent, a modular, cross-environment harness for Robot-Use Agents, with paper, code and docs public.
Key results: on the 42-task RoboDojo benchmark, Qwen3.8-Flash-Next scored 31.39, beating GPT-6 Astra's public 28.97 — at $1.09 per episode vs Astra's $21.29 (6.6% of the cost); a full 252-episode eval costs <$300 vs >$5,000. On 19 real-robot tasks, Qwen scored 53.50 with zero real-world training data, beating fine-tuned π0.5 (47.75); Astra hit 80.50. The same harness drove a Franka arm and Sharpa dexterous hand through grasping and button-pressing tasks.
Design principles: unified tool definitions across sim and real, plug-and-play modularity, strictly real-world-available information (no depth/CAD/hidden annotations), and background parallel execution of actions. Ablations: removing depth barely hurts Astra (55.83→55.04) but drops Qwen to 31.39; removing calibration too crashes Qwen to 9.13. Disabling the skill library retains 96.6% of Qwen's score.
Failure-trajectory analysis identifies four weak spots (scene understanding, cross-view localization, geometric computation, physical plausibility of actions), and RoboDojo scores correlate with embodied-reasoning and multimodal benchmarks — hints for future training. Modeled after mini-swe-agent, the project aims to export interaction trajectories as SFT/RL data, enabling an open-weight learning loop for robots.
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
- Open-weight LLMs control robots zero-shot: Qwen scores 31.39 on RoboDojo, beating GPT-6 Astra — pabbeel · 2026-11-04