The field of robot learning is moving towards addressing the challenges of learning from imbalanced and limited datasets. Researchers are exploring methods to rebalance datasets, reduce shortcut learning, and improve generalization capabilities of generalist robot policies. Novel approaches include analogical reasoning, variational bottlenecks, and physical autoregressive models. Notable papers include:
- Towards Balanced Behavior Cloning from Imbalanced Datasets, which introduces a meta-gradient rebalancing algorithm to address the limitations of existing approaches.
- AR-VRM: Imitating Human Motions for Visual Robot Manipulation with Analogical Reasoning, which proposes a keypoint Vision-Language Model pretraining scheme to learn human action knowledge.
- Masquerade: Learning from In-the-wild Human Videos using Data-Editing, which edits in-the-wild egocentric human videos to bridge the visual embodiment gap between humans and robots.