Motion Policy Networks

TL;DR

Motion Policy Networks (MπNets) address collision-free motion in unknown environments by learning an end-to-end policy that maps a segmented point cloud and a robot configuration to a normalized joint-space displacement, enabling real-time, reactive motion without explicit scene models. Trained on a massive synthetic dataset of over $3.0$ million problems across more than $5 imes 10^5$ environments, MπNets combine a two-encoder architecture with geometric behavior cloning and collision losses, and are trained to roll out recursively in closed loop. The approach achieves substantially faster planning than traditional global planners while maintaining high success rates, and it outperforms prior neural planners and local control policies on challenging, partially observed, and dynamic scenarios, with demonstrated sim-to-real transfer on a 7-DOF robot. Limitations include reliance on expert quality and potential generalization gaps, suggesting future work with DAgger, domain adaptation, and learned collision checking for safer deployment. Overall, MπNets offer a scalable, perception-driven alternative to traditional planning pipelines, enabling practical, real-time manipulation in unknown environments.

Abstract

Collision-free motion generation in unknown environments is a core building block for robot manipulation. Generating such motions is challenging due to multiple objectives; not only should the solutions be optimal, the motion generator itself must be fast enough for real-time performance and reliable enough for practical deployment. A wide variety of methods have been proposed ranging from local controllers to global planners, often being combined to offset their shortcomings. We present an end-to-end neural model called Motion Policy Networks (M$π$Nets) to generate collision-free, smooth motion from just a single depth camera observation. M$π$Nets are trained on over 3 million motion planning problems in over 500,000 environments. Our experiments show that M$π$Nets are significantly faster than global planners while exhibiting the reactivity needed to deal with dynamic scenes. They are 46% better than prior neural planners and more robust than local control policies. Despite being only trained in simulation, M$π$Nets transfer well to the real robot with noisy partial point clouds. Code and data are publicly available at https://mpinets.github.io.

Figures (5)

• Figure 1: M$\pi$Nets are trained on a large dataset of synthetic demonstrations (left) and can solve complex motion planning problems using raw point cloud observations (right).
• Figure 2: M$\pi$Nets encodes state as a normalized robot configuration and segmented point cloud with three classes for the robot, the obstacles, and the target. The policy outputs a displacement in normalized joint space, which can then be applied to the input before unnormalizing to get $q_{t+1}$.
• Figure 3: M$\pi$Nets is trained with a dataset consisting of solutions to $3.27$ million unique planning problems across over 575000.0 unique, procedurally generated environments.
• Figure 4: M$\pi$Nets performance continues to increase with more training data, while MPNets performance stays relatively constant
• Figure 5: After injecting Gaussian noise into the point clouds, M$\pi$Nets performance stays fairly constant up until $\sigma=3\cm$ when success rate is 89.28%.