Mitigate catastrophic forgetting in generalist robotic manipulation policies

Develop training strategies and model designs that prevent catastrophic forgetting when generalist robotic manipulation policies are trained across many tasks, ensuring retention of previously learned skills while acquiring new ones.

Background

The paper studies Large Behavior Models (LBMs)—multitask visuomotor policies trained on large, heterogeneous robot datasets—and evaluates their benefits over single-task models. While demonstrating gains from multitask pretraining, the authors note that catastrophic forgetting persists as an open research problem for generalist policies.

Addressing forgetting is critical for practical deployment of generalist policies that must continually incorporate new tasks and data without erasing prior capabilities.