Summary
This paper presents the first systematic study of continual learning for VLA models under realistic real-world conditions. The authors construct a real-world continual learning dataset spanning four sequential manipulation tasks (rigid pick-and-place, contact-rich pressing, deformable folding, and multi-object long-horizon). The key finding is that VLA models suffer significant catastrophic forgetting when continually learning from heterogeneous real-world demonstrations. Experience replay is evaluated and key implementation factors governing its success are identified.
Key Contributions
- First real-world continual learning dataset for VLAs covering four diverse sequential manipulation tasks
- Empirical demonstration of significant catastrophic forgetting in VLA models under realistic continual learning conditions
- Systematic evaluation of experience replay as a mitigation strategy with analysis of critical implementation factors
- Practical guidelines for deploying VLA models in settings where new skills must be acquired without resetting knowledge
Significance
Closes a major gap between lab VLA evaluations (fixed task distributions) and real-world deployment (sequential skill acquisition), providing the community with rigorous empirical baselines and a reusable benchmark dataset.