RoboACE: Adversarial Co-Evolution for Robot Self-Improvement
RoboACE: Adversarial Co-Evolution for Robot Self-Improvement
This research presents RoboACE, a framework enabling robot agents to self-improve by jointly evolving curriculum and execution harnesses around frozen foundation models. The system alternates updates between proposing informative challenges based on accumulated capability evidence and responding with reusable revisions that expand execution capabilities. Starting from a compact set of LIBERO practice conditions, the resulting frozen deployment achieves 74.0% success on LIBERO-PRO, exceeding the initial deployment by 15.0 percentage points and direct execution by its VLA controller by 13.5 points. Case studies document changes to evidence-guided practice selection, drawer-skill grounding, and controller intervention.
The paper introduces RoboACE, a framework for adversarial co-evolution that jointly updates curriculum and execution harnesses around frozen foundation models to improve robot skills on the LIBERO benchmark.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
isions. Starting from a compact set of LIBERO practice conditions, the resulting frozen deployment outperforms the evaluated controllers and agent systems across four LIBERO-family panels. On LIBERO-PRO, it reaches 74.0% success, exceeding the initial deployment by 15.0 percentage points and direct execution by its VLA controller by 13.5 points. Case studies document changes to evidence-guided practice selection, dra
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org