Runjia Tan, Yuang Tu, Yujie Yan, Lan Yu, Xuesong Tian, Chen Lv
FIERCE is a reinforcement learning framework that refines generalist robot policies into compact specialists using progress-failure feedback. It achieves this through a unified task-adaptive evaluator and joint supervision from progress and preference labels. FIERCE enables efficient refinement without dedicated simulators or manual rewards.
FIERCE has practical significance as it enables robots to learn from limited physical interaction, making it suitable for real-world applications where simulation or manual annotation may not be feasible. This can lead to more efficient and effective robot learning in various domains. The framework's ability to refine generalist policies into specialists can also reduce deployment costs.
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.