Researchers at Stanford have introduced RoboFuME, a system for autonomous and effective real-world robot learning. It addresses challenges in fine-tuning robot policies and uses a language-conditioned, offline reinforcement learning multitask strategy. The system leverages calibrated offline reinforcement learning techniques and a reward predictor to reduce the need for human input during online fine-tuning. It also utilizes vision-language models to refine pre-trained representations and create surrogate rewards. The system has shown notable advantages in various real-world tasks, offering improvements over offline-only techniques.
1 Impression