Democratizing Robot-Use Agents: A Unified Harness with Open Models
Democratizing Robot-Use Agents: A Unified Harness with Open Models
This paper introduces uni-robot-agent, an open-weight harness for robot-use agents that supports further training without relying on proprietary models. Using Qwen3.8-Flash-Next out-of-the-box, the system achieves a RoboDojo Score of 31.39, surpassing the published GPT-6 Astra-RoboProbe result of 28.97 and performing competitively on 19 real-world tasks. The authors analyze model generations, families, and sizes to link effective robot use to four general capabilities, suggesting future training should develop them jointly.
The authors introduce uni-robot-agent, an open-weight harness designed to enable further training of robot-use agents without relying on proprietary models; the reported RoboDojo Score of 31.39 is achieved by Qwen3.8-Flash-Next under out-of-the-box conditions and is compared against a published GPT-6 Astra-RoboProbe result of 28.97.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
RoboDojo Score
31.39
View original evidence
ash-Next achieves a RoboDojo Score of 31.39, surpassing the published GPT-6 Astra-RoboProbe result of 28.97 and many VLA/WAM baselines, and performs competitively on 19 real-world tasks. Our analysis across model generations, families, and sizes links effective robot use to four general capabilities, suggesting that future RUA training should develop them jointly.
Open source S3GPT-6 Astra-RoboProbe RoboDojo Score
28.97
View original evidence
ash-Next achieves a RoboDojo Score of 31.39, surpassing the published GPT-6 Astra-RoboProbe result of 28.97 and many VLA/WAM baselines, and performs competitively on 19 real-world tasks. Our analysis across model generations, families, and sizes links effective robot use to four general capabilities, suggesting that future RUA training should develop them jointly.
Open source S3RoboDojo Score (GPT-6 Astra-RoboProbe)
28.97
View original evidence
ash-Next achieves a RoboDojo Score of 31.39, surpassing the published GPT-6 Astra-RoboProbe result of 28.97 and many VLA/WAM baselines, and performs competitively on 19 real-world tasks. Our analysis across model generations, families, and sizes links effective robot use to four general capabilities, suggesting that future RUA training should develop them jointly.
Open source S3RoboDojo Score (uni-robot-agent)
31.39
View original evidence
ash-Next achieves a RoboDojo Score of 31.39, surpassing the published GPT-6 Astra-RoboProbe result of 28.97 and many VLA/WAM baselines, and performs competitively on 19 real-world tasks. Our analysis across model generations, families, and sizes links effective robot use to four general capabilities, suggesting that future RUA training should develop them jointly.
Open source S3
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ash-Next achieves a RoboDojo Score of 31.39, surpassing the published GPT-6 Astra-RoboProbe result of 28.97 and many VLA/WAM baselines, and performs competitively on 19 real-world tasks. Our analysis across model generations, families, and sizes links effective robot use to four general capabilities, suggesting that future RUA training should develop them jointly.
Open source S3
Source：arXiv Robotics — research abstracts · arxiv.org