Open Harness Advances Robot-Use Agent Research

Researchers introduced uni-robot-agent, a unified harness for robot-use agents built with open-weight models, to make further training research more accessible. The paper was published on Nov 1 and submitted to Hugging Face on Nov 3. Robot-use agents have shown zero-shot manipulation (handling tasks without task-specific examples), but reliance on frontier proprietary models has limited further training research.

Without privileged inputs, the out-of-the-box Qwen3.8-Flash-Next scored 31.39 on RoboDojo, above the published GPT-6 Astra-RoboProbe result of 28.97 and many VLA and WAM baselines. It also performed competitively across 19 real-world tasks. The researchers say the unified harness supports further training of open-weight robot-use agents.

The team analyzed models across generations, families, and sizes and linked effective robot use to four general capabilities. They suggest future robot-use-agent training should develop those capabilities jointly. The source does not name the four capabilities or give further task-level scores, so the reported results are limited to the overall RoboDojo score and performance across 19 tasks.

Researchers introduced uni-robot-agent, a unified harness for robot-use agents built with open-weight models, to make further training research more accessible. The paper was published on Nov 1 and submitted to Hugging Face on Nov 3. Robot-use agents have shown zero-shot manipulation (handling tasks without task-specific examples), but reliance on frontier proprietary models has limited further training research.

Without privileged inputs, the out-of-the-box Qwen3.8-Flash-Next scored 31.39 on RoboDojo, above the published GPT-6 Astra-RoboProbe result of 28.97 and many VLA and WAM baselines. It also performed competitively across 19 real-world tasks. The researchers say the unified harness supports further training of open-weight robot-use agents.

The team analyzed models across generations, families, and sizes and linked effective robot use to four general capabilities. They suggest future robot-use-agent training should develop those capabilities jointly. The source does not name the four capabilities or give further task-level scores, so the reported results are limited to the overall RoboDojo score and performance across 19 tasks.