Never Forget & Never Stop Learning
A large-scale real-world dataset designed for continual robot learning, capturing diverse tasks and deployment conditions for lifelong policy improvement.
The first empirical study of real-world continual VLA learning. We find VLA models suffer significant catastrophic forgetting on heterogeneous real-world tasks, and identify key implementation factors that govern the success of experience replay for mitigating forgetting.
We explore continual learning for vision transformers so visual representations can adapt to new data streams without erasing knowledge acquired on earlier tasks.
We present a teacher–student framework that decouples continual RL into distributed single-task teacher training and continual distillation into a central generalist model, mitigating forgetting while scaling to sequential tasks.