XFlowMP: Task-Conditioned Motion Fields for Generative Robot Planning with SchröDinger Bridges

Abstract

Generative robotic motion planning requires synthesizing smooth, collision-free trajectories that remain feasible across diverse tasks and dynamic constraints. Prior planners, both traditional and generative, often struggle to couple high-level semantics with low-level constraints, especially the link between task configurations and motion controllability. We present XFlowMP, a task-conditioned generative motion planner that models robot trajectory evolution as entropic flows bridging stochastic noise and expert demonstrations via Schrödinger bridges, conditioned on the inquiry task configuration. Our method couples Schrödinger bridges as a conditional flow matching with a score function to learn motion fields with high-order dynamics while encoding start-goal configurations, generating collision-free and dynamically-feasible motions. In evaluations, XFlowMP achieves up to 53.79% lower maximum mean discrepancy, 36.36% smoother motions, and 39.88% lower energy consumption over the next-best baseline on the RobotPointMass benchmark, while reducing short-horizon planning time by 11.72%. On long-horizon motions in LASA Handwriting, it maintains 1.26% lower discrepancy, 3.96% smoother, and 31.97% lower energy. We further demonstrate practicality on the Kinova Gen3 manipulator, confirming robustness in real-world settings. Our source code is available at: https://github.com/mkhangg/xflow.mp.