World action models (WAMs) built on pretrained video generation models adapt quickly to new robotic tasks. However, when fine-tuned sequentially on new tasks, they tend to forget previously learned ones, a problem the paper calls catastrophic forgetting.

Experience replay is a common remedy, but it requires storing past data, which can be impractical in real-world settings. A new arXiv paper proposes a data-free alternative using rectified generative replay, which generates synthetic experiences from the model itself rather than relying on stored samples.

The abstract suggests this approach could mitigate forgetting while keeping the model adaptable to new tasks. The paper is available on arXiv and presents a method that may offer a practical path for continual learning in robotics without the overhead of data storage.