MEKAWIRE
IntermediatearXiv

JustRUA: A Controlled Study of GPT-6 Robot-Use Agents

Zefan Wang, Yufan Dang, Xuanwei Shi, Mingqi Gao, Cheng Yin, SikYuen Tam, Hanyu Liu, Yasheng Wang, Yaxi Lu, Songyang Zhang, Xin Cong, Yuan Yao
In 30 seconds

This preprint presents JustRUA, a software harness that lets GPT-6 Astra command simulated robot arms without a learned control policy. On 111 episodes covering 42 RoboDojo tasks, its score reached 50.3 versus 29.7 for the official L3 harness. The results show how observation history, motion interruption and spatial information can materially change a general-purpose model's manipulation performance.

Research question

Which changes to the harness, action interface, experience reuse and scene geometry improve GPT-6 Astra's robot control under a shared evaluation protocol?

Problem

A model can choose a sensible task plan yet miss objects, misjudge placement or let an unstable motion continue before receiving another observation. Published robot-use agents also change multiple components and evaluation conditions together, making their reported gains difficult to attribute.

Previous approach

The official L3 Inspect EEF harness retains only the latest two image observations and executes complete CuRobo plans toward absolute end-effector targets. Systems including RoboDawn, RoboHarness and ORA add demonstrations, perception or other interfaces, but their differing protocols obscure component-level benefits. Learned VLA policies are another control family; JustRUA instead examines a fixed interface without an intervening learned policy.

New approach

JustRUA preserves episode images, actions and feedback in a Codex session, interrupts each motion after at most 20 control steps, and retains absolute world-frame targets with the original CuRobo planner. Observations add compact gripper measurements, raw Depth Pro estimates from RGB, and camera calibration. GPT-6 Astra can execute code to calculate 3D targets from image pixels. The final configuration also removes the 100-decision limit while retaining the benchmark's 500-control-step limit. It adds no learned policy, cross-episode experience or task-specific tuning beyond the inherited task recipes. The authors report releasing code, the evaluation set and replays.

Results

Scores measure task progress on a 0 to 100 scale, not just full success. Across 111 simulated episodes covering 42 tasks, full history plus the 20-step cap scored 38.0; adding compact readings scored 39.8, predicted depth 42.4, and coordinate-solving code together with removal of the decision cap 50.3. The final gain over the official harness's subset score of 29.7 was 20.6 points; final success rate was 43.3%. The official harness scored 29.0 on the separate full protocol of 2,100 episodes. In the same 42 first-round episodes, predicted depth reduced empty grasps from 41 with compact readings to 16. Cached-input cost lower bounds per first-round episode were $4.94 for the default testbed, $7.74 with predicted depth and $12.53 for JustRUA. Code and decision-budget effects were not isolated.

Limitations

All experiments use simulation, one benchmark and one model, with one evaluation-set run per configuration. Components were tested along a sequential development path rather than in every combination. The subset was selected using the authors' default testbed and differs from the official layout distribution, so these are not leaderboard scores. The final step combines code execution with a larger decision budget. Behaviour counts are evidence-based lower bounds. Experience-reuse tests use evaluation tasks and only five improvement rounds, with one practice run per task. Remaining failures include unsuitable plans, collisions and slipping grasps. Costs assume cached inputs; real-robot performance and deployment latency are not reported in the paper.

Industry impact

Robot-arm manufacturers and manipulation software suppliers could use these findings now to guide simulation experiments: preserve useful observations, shorten unchecked motions and supply spatial evidence efficiently. Applying the approach to hardware products first requires real-robot validation, reliable sensing, collision handling and acceptable inference cost and latency. A commercial deployment timeline is not reported in the paper.

The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.