The MeanFlow Tames Policy Learning in 1-step for Robotic Manipulation (MP1) is a new robotic learning framework proposed by a research team from Peking University on July 14, 2025. It combines 3D point cloud input with the MeanFlow framework and can directly generate action trajectories within a single network function evaluation.

MP1 utilizes the MeanFlow architecture to complete policy generation in a single forward propagation, balancing speed, accuracy, and generalization, breaking through the performance bottlenecks of both diffusion and flow models. Furthermore, by innovatively combining MeanFlow representation, CFG conditional control, and dispersive loss, it achieves a one-step approach to generating robotic manipulation policies that is fast, accurate, and stable.