Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning

Continual learning is a long-standing challenge in robot policy learning, where a policy must acquire new skills over time without catastrophically forgetting previously learned ones. While prior work has extensively studied continual learning in relatively small behavior cloning (BC) policy models trained from scratch, its behavior in modern large-scale pretrained Vision-Language-Action (VLA) models remains underexplored. In this work, we find that pretrained VLAs are remarkably resistant to forgetting compared with smaller policy models trained from scratch. Simple Experience Replay (ER) works surprisingly well on VLAs, sometimes achieving zero forgetting even with a small replay data size. Our analysis reveals that pretraining plays a critical role in downstream continual learning performance: large pretrained models mitigate forgetting with a small replay buffer size while maintaining strong forward learning capabilities. Furthermore, we find that VLAs can retain relevant knowledge from prior tasks despite performance degradation during learning new tasks. This knowledge retention enables rapid recovery of seemingly forgotten skills through finetuning. Together, these insights imply that large-scale pretraining fundamentally changes the dynamics of continual learning, enabling models to continually acquire new skills over time with simple replay.

Robots are increasingly expected to learn many tasks over time, but a major challenge is that learning a new task can make them worse at tasks they previously knew. This problem is often called “forgetting” and is one of the key obstacles to building robots that can continue improving after deployment. In this work, we study whether large pretrained vision-language-action models for robotics suffer from this problem when learning tasks sequentially. Surprisingly, we find that these pretrained models are much more resistant to forgetting than smaller robot policies trained from scratch. With a simple method that occasionally reuses a small amount of data from old tasks, models such as Pi0 and GR00T can often learn new tasks without losing performance on earlier ones. In some cases, learning new tasks even improves performance on old tasks. We also find that when performance drops, the model has not always truly lost the old knowledge; it can often recover quickly with only a small amount of additional training. These results suggest that pretrained robot models may provide a stronger foundation for continual learning than previously expected, bringing us closer to robots that can keep learning new skills over time without constantly starting over.