JustRUA: A Controlled Study of GPT-6 Robot-Use Agents
A stripped-down harness lifted the RoboDojo simulation score from 29.7 to 50.3 without task-specific tuning.
The study tests harness design, action interfaces, experience reuse and scene geometry within one framework to isolate their effects. Keeping the full episode history and stopping each motion early delivered substantial gains. Scene geometry helped most, even when depth was predicted from RGB images alone. Reusing experience changed the model’s behavior but did not offer an unqualified benefit.
HF Daily Papers' note
The study tests harness design, action interfaces, experience reuse and scene geometry within one framework to isolate their effects. Keeping the full episode history and stopping each motion early delivered substantial gains. Scene geometry helped most, even when depth was predicted from RGB images alone. Reusing experience changed the model’s behavior but did not offer an unqualified benefit.
HF Daily Papers' note
score 5