FlowMPC: Improving Flow Matching policies with World Models
Researchers have developed FlowMPC, a new framework that enhances the performance of Flow Matching (FM) policies in multimodal action spaces. By integrating a learned world model with an imitation-learned FM policy, FlowMPC enables Model Predictive Path Integral (MPPI) planning for improved test-time actions. Experiments on ManiSkill manipulation tasks, specifically PickCube and PickSingleYCB, demonstrated that FlowMPC outperformed FM policies alone, showing notable gains in end-of-episode success rates. This approach suggests that world-model-based planning can effectively complement FM imitation policies without altering their training objective. AI
IMPACT Enhances imitation learning for robotics by combining flow matching with world models for improved planning.