Perceptual Local Collision-Free Motion Generation for Manipulators

Collision-free motion planning has a well-established research history, yet the majority of studies have been centered around Cartesian space and assume the exact positions and shapes of the obstacles in the environment are known. In this letter, we introduce a perceptual local reactive motion controller for manipulators that navigates to desired configurations while avoiding collisions with both static and dynamic obstacles, as well as self-collisions. We formulate collision-free motion generation as a constrained quadratic programming problem, where collision avoidance is achieved by linearizing minimum distance constraints. We have integrated obstacle velocity into the planning algorithm to ensure reliable collision avoidance in the presence of dynamic obstacles. To enable fast collision distance queries in high-dimensional configuration space, we approximate implicit signed distance functions using neural networks and leverage forward kinematics to eliminate non-linear mappings between configuration and Cartesian spaces. This approach improves both computational efficiency and fitting accuracy. The gradients of the distance function with respect to joint states are provided for collision constraints using the derived analytical gradient. Additionally, self-collision avoidance is achieved using a similar method. The proposed planner typically solves within 10 milliseconds, even in scenarios with 10,000 points. Both simulation and physical experiments validate the effectiveness of our approach.