BORA: Bridging Offline Reinforcement Learning and Online Residual Adaptation for Real-World Dexterous VLA Models
BORA is a framework that takes offline reinforcement learning skills from Vision-Language Models and adapts them to real-world robot execution. It uses an action-conditioned critic during offline training and then applies a lightweight, human-in-the-loop residual adaptation online to correct physical errors. This method improves success rates by grounding value estimation in physical actions rather than just visual context.