JustRUA harness lifts GPT-6 Astra on RoboDojo, 29.7 to 50.3

GPT-6 Astra scores 50.3 on the RoboDojo simulation benchmark when wrapped in a lean harness, up from 29.7 on the official L3 setup, with no learned policy, no self-improvement loop and no task-specific tuning.

That is the finding of JustRUA, a controlled-ablation paper by Zefan Wang and colleagues posted to arXiv on November 11. The authors pull apart four design axes: the harness around the model, the action interface, the reuse of experience, and the scene geometry given to the model. They test them in one framework rather than stacked together.

'A lean harness that keeps the full episode history and stops each motion early goes a long way,' the paper reports. Scene geometry helps most, 'even when depth is predicted from RGB alone.' Experience reuse changes how the model acts but is, in the authors' words, 'not a free lunch.'

The framing matters because, as the abstract puts it, '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.' JustRUA starts from the official L3 harness and keeps only what matters most. The abstract reports no per-task numbers.