Single Swept Volume Reconstruction by Signed Distance Function Learning: A feasibility study based on implicit geometric regularization

Abstract

In automated production using collaborative robots in a manufacturing cell, a crucial aspect to ensure the safety of workers and robots in human-robot interaction is to avoid collisions or detect unsafe situations. When detecting collisions, one approach is the use of swept volume to identify safe protective space for the operation or maintenance. To reduce the cost of collision checking due to the complexity of swept volume geometries, we propose to use Implicit Geometric Regularization (IGR) to train the signed distance function of the precomputed swept volume as a continuous collision detector during task trajectory execution. We validate the accuracy of the reconstruction given by Hausdorff distance and memory budgets over three trajectory instances followed by planar 2-link and PUMA-like fixed-base robot arms with polyhedral links and rotary joints. The simulation results and the error analysis of test scenarios suggest that it is feasible to model the signed distance function of a precomputed swept volume by IGR.