Researchers presenting RoboACE propose a way for robot agents to improve by revising both the challenges they practice on and the routines they use to carry them out. The framework works around frozen foundation models, alternating between updates to a practice curriculum and updates to an execution harness.
In the researchers’ evaluations, the resulting frozen deployment achieved 74.0% success on LIBERO-PRO. That was 15 percentage points above the initial deployment and 13.5 points above direct execution by its vision-language-action controller.
Why it matters
For robotics researchers evaluating ways to improve robot skills without changing the underlying model, RoboACE reports gains from revising practice selection and execution routines; whether those gains carry over beyond the tested LIBERO-family tasks remains unknown.
