Tsinghua, Berkeley Harness Lets Qwen3.8 Edge Fine-Tuned π₀.₅
TL;DR
- Qwen3.8-Flash-Next scored 31.39 on the RoboDojo simulation benchmark inside the uni-robot-agent harness, above the GPT-6 Astra–RoboProbe published 28.97.
- On 19 real-world tabletop tasks the open stack scored 49.50, with the paper listing fine-tuned π₀.₅ at 47.75.
- The authors link effective robot use to four general capabilities and argue future RUA training should develop them jointly.
Qwen3.8-Flash-Next, slotted into a new open-weight agent harness, scored 31.39 on the RoboDojo simulation benchmark. GPT-6 Astra–RoboProbe's published number on the same test is 28.97. The harness, called uni-robot-agent, comes from a team at Tsinghua University and UC Berkeley.
In the paper posted on Hugging Face, the authors frame the project as a push against researchers' dependence on closed frontier models. 'To democratize RUA research, we introduce uni-robot-agent, a unified harness for competitive open-weight robot use that supports further training,' they write.
On 19 real-world tasks the open stack held up. Qwen3.8-Flash-Next scored 49.50; the paper lists π₀.₅, labelled fine-tuned, at 47.75. The authors tie strong robot-use performance to four general capabilities and argue future training should develop them jointly.
The paper also flags a cost gap at inference: $1.09 per episode for Qwen, $21.29 for Astra. It lands in the middle of a steady run of robotics coverage we have been tracking this week.
Originally reported by huggingface.co
Read the original article →Original headline: Tsinghua and Berkeley Uni-Robot-Agent Harness Matches Fine-Tuned pi 0.5 on 19 Real-World Tabletop Tasks