FlowMPC: Improving Flow Matching policies with World Models

A new framework called FlowMPC combines Flow Matching imitation policies with learned world models to improve robot manipulation at test time, according to research posted to arXiv. The approach uses model-based planning to boost end-of-episode success without altering the original training objective. Flow Matching has gained traction as a method for behavior cloning in multimodal action spaces, but its policies are not trained to directly maximize expected return [1]. That gap leaves room for improvement during deployment. The FlowMPC framework addresses this by pairing an imitation-learned Flow Matching policy with a learned world model, then applying Model Predictive Path Integral planning over candidate action sequences proposed by the policy [2]. The architecture builds on TD-MPC2, a model-based reinforcement learning system introduced by Hansen et al. in 2024 [2]. The work was evaluated on two manipulation benchmarks from the ManiSkill suite, introduced by Tao et al. in 2025: PickCube and PickSingleYCB [2]. In both tasks, adding the world model improved performance over the Flow Matching policy alone. The gains were most pronounced in end-of-episode success rates, suggesting that planning with a learned dynamics model helps the policy recover from states where imitation alone falls short [2]. Unlike prior efforts that modify the training objective of Flow Matching policies, FlowMPC leaves the original behavior-cloning loss intact. The world model and planner operate only at test time, making the approach compatible with existing pre-trained Flow Matching checkpoints [2]. The paper does not report results on more complex, contact-rich tasks, and the author notes that scaling to higher-dimensional action spaces remains an open question. The preprint was submitted to arXiv on June 15, 2026, under the Machine Learning category [1]. The research bundle includes references to Jiang et al.'s 2025 work on Flow Matching for behavior cloning and Hansen et al.'s 2024 TD-MPC2 paper, situating FlowMPC within a broader line of work that seeks to combine the strengths of imitation learning and model-based planning [2].