Qwen-powered uni-robot-agent hits 31.39 on RoboDojo, beating GPT-6 Astra at 1/19.5 the cost
dotey · x · 2026-11-05
- uni-robot-agent is a unified harness for Robot-Use Agents (RUAs), combining perception, motion, skills and termination logic, with adapters for RoboDojo/Libero/RoboLab sims and real robots (ARX X5, Franka, Sharpa) — a first step toward RUA training and local deployment.
- Tested with locally-run Qwen3.8-Flash-Next: zero-shot RoboDojo score jumped from 1.90 to 31.39, surpassing the RoboDojo Team's official GPT-6 Astra result of 28.97, at roughly 1/19.5 the cost.
- On 19 real-world robot tasks it scored 53.50, beating fine-tuned π0.5 (47.75) by 5.75 points and zero-shot Dream-1 (43.50) by 10 points.
More from Embodied
- NYU and Meta FAIR introduce ID-WM, a hierarchical world model beating baselines by 38.5% — ylecun · 2026-11-05
- mjorbit: open-source space robotics simulator runs thousands of parallel worlds on one GPU — ChongZzZhang · 2026-11-05
- Centauro-style robot with humanoid torso and wheeled legs goes viral — chrisalbon · 2026-11-05
- Scoble to bring a rentable booth robot to CES after seeing the stunt work thousands of times — Scobleizer · 2026-11-05
- USC paper shows VLAs can keep the same rationale while adversarial perturbations change their actions — USC · 2026-11-05
- 11,398 real-world evals: GPT-6 Astra stands out in 5-VLA policy comparison — DJiafei · 2026-11-05