Dream to Remember: Continual Steering of Generalist Robot Policies with World Models
The paper’s claim is that a frozen world model can help a robot policy learn new tasks without erasing old ones.
Dream to Remember trains only a steering module around a frozen generalist robot policy, using imagined rollouts to rehearse earlier skills. On CALVIN, the method improved eight tasks in sequence in imagination, lifting average success from 63.7% to 74.8%, while LoRA fine-tuning fell to 11.7%. The authors also report results on LIBERO-90 and a real Franka robot, where D2R retained earlier gains while adding new tasks. ArXiv · AI/CL/LG's note
Dream to Remember trains only a steering module around a frozen generalist robot policy, using imagined rollouts to rehearse earlier skills. On CALVIN, the method improved eight tasks in sequence in imagination, lifting average success from 63.7% to 74.8%, while LoRA fine-tuning fell to 11.7%. The authors also report results on LIBERO-90 and a real Franka robot, where D2R retained earlier gains while adding new tasks. ArXiv · AI/CL/LG's note
score 5