Fast Collision Detection for Robot Manipulator Path: an Approach Based on Implicit Neural Representation of Multiple Swept Volumes

Abstract

Detecting the collision to ensure safety is required in automated production using robots, specifically when the robot operates in environments with obstacles or human-robot collaboration. One approach for detecting collisions is using the swept volume (SV) to identify a safe protective space for operation. To reduce the computational time and memory of collision checks due to the complexity of SV geometries, we propose to first construct an accurate and reliable signed distance function (SDF) via deep learning from raw point clouds of a collection of precomputed SVs as an implicit neural representation of the whole workspace, which only requires to store a set of parameters. We leverage the learned SDF as a continuous collision detector in RRT path planning for a planar 2R manipulator in a cluttered environment. A 24% reduction in the computational time at the cost of a 5% reduction in detection accuracy compared to exact collision detection with a real SV is observed for trading accuracy for faster path computation using our approach. In a simulation in a simple dynamic environment, GPU-accelerated implementation of DRRT with SDF achieves 4x-62x faster path computation. Discontinuity of signed distance of multi-SVs that results in learning accuracy degradation is also discussed.