The paper introduces uni-robot-agent, a unified harness for open-weight robot-use agents that is built to support further training.
Headline result: Without privileged inputs, an out-of-the-box Qwen3.8-Flash-Next scores 31.39 on the RoboDojo Score. The paper says that beats the 28.97 published for GPT-6 Astra-RoboProbe and many VLA and WAM baselines.
Real robots: The authors say it performs competitively on 19 real-world tasks.
Capabilities: Across model generations, families and sizes, the analysis links effective robot use to four general capabilities.
Suggestion: The authors argue future robot-use agent training should develop those four capabilities jointly.
Sign in to comment.
TL;DR of the paper, generated automatically.
The paper introduces uni-robot-agent, a unified harness for open-weight robot-use agents that is built to support further training.