RECALL: Recovery Experience Collection for Active Lifelong Learning in Vision-Language-Action Models

The paper presents a novel active continual learning paradigm for Vision-Language-Action (VLA) models, emphasizing the benefits of uncertainty-guided data collection over traditional passive imitation learning. Key findings include improved fine-tuning efficiency when using actively collected recovery data, though this approach risks catastrophic forgetting if not managed properly. Techniques such as replay-based data mixing and elastic weight consolidation are evaluated, highlighting the trade-offs between adapting to new data and retaining previously learned behaviors, which is crucial for practitioners developing robust VLA systems.