JustRUA raises RoboDojo robot-agent score from 29.7 to 50.3
Original title: JustRUA: A Controlled Study of GPT-6 Robot-Use Agents
A lean harness around GPT-6 Astra, with no learned policy, lifts the simulation benchmark by about 20 points.
JustRUA isolates design choices for GPT-6 robot-use agents that drive a robot arm through a fixed harness on RoboDojo. The authors compare the harness, action interface, experience reuse and scene geometry inside one framework. A lean harness that keeps the full episode history and stops each motion early accounts for much of the gain. Scene geometry helps most, even when depth is predicted from RGB alone. Starting from the official L3 harness and without a learned policy, self-improvement loop or task-specific tuning, the score rises from 29.7 to 50.3.
Unconfirmed: The 50.3 score is the authors' RoboDojo simulation result, not an independent evaluation.
- The paper starts from the official L3 harness and keeps only the strongest pieces.
- Experience reuse changes how the model acts but is not a free lunch.
- Scene geometry helps most, including depth predicted from RGB alone.
- No learned policy, self-improvement loop or task-specific tuning is used.
From the original article
1 min read at huggingface.co
GPT-6 Astra can drive a robot arm directly through a fixed harness, and scores of such robot-use agents on RoboDojo have risen quickly since. Reported systems differ in harness and test subset and often change several designs at once, so it is hard to tell which choices produce the gains.
By Zefan Wang, Yufan Dang, Xuanwei Shi · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.