JustRUA isolates the design choices behind GPT-6 robot agents
The JustRUA preprint presents a controlled study of GPT-6 Astra agents operating a robot arm, testing harness design, action interface, experience reuse, and scene geometry within one framework. Its simplified setup raises the reported RoboDojo simulation score from 29.7 to 50.3 without a learned policy, self-improvement loop, or task-specific tuning. The authors find that a lean harness with full episode history and early motion termination is especially effective, while scene geometry provides the largest benefit among the tested designs.
By testing several robot-agent design choices within a unified framework, JustRUA offers a clearer account of which configuration changes contribute to reported RoboDojo performance gains.
Computer Science > Robotics
arXiv:2611.16009 (cs)
[Submitted on 11 Nov 2026]
Title:JustRUA: A Controlled Study of GPT-6 Robot-Use Agents
Authors:Zefan Wang, Yufan Dang, Xuanwei Shi, Mingqi Gao, Cheng Yin, SikYuen Tam, Hanyu Liu, Yasheng Wang, Yaxi Lu, Songyang Zhang, Xin Cong, Yuan Yao
View a PDF of the paper titled JustRUA: A Controlled Study of GPT-6 Robot-Use Agents, by Zefan Wang and 11 other authors