Teaching robot policies without new demonstrations: interview with Jiahui Zhang and Jesse Zhang - Robohub

The article discusses the ReWiND framework introduced by Jiahui Zhang, Jesse Zhang, and colleagues in their CoRL 2025 paper, which enables robots to learn manipulation policies for novel, language-specified tasks without requiring new task-specific demonstrations. ReWiND operates in three stages: first, it learns a dense reward function from a small set of demonstrations in the deployment environment by predicting per-frame progress toward task completion. A novel video rewind augmentation technique is used to synthetically generate sequences simulating both progress and failure, improving the reward model’s accuracy and generalization. Second, the framework pre-trains a policy offline using these dense rewards relabeled on the demonstration data. Finally, the pre-trained policy is fine-tuned online in the deployment environment on unseen tasks, using the frozen reward function to provide feedback without additional demonstrations. The researchers evaluated ReWiND in both simulated (MetaWorld) and real-world (Koch) environments, focusing on the reward model’s generalization and the policy