DistFormer: A High-Accuracy Transformer-Based Collision Distance Estimator for Robotic Arms
Yue Cao, Jiakang Zhou, Feng Cao, Feng Shu · 2023
Collision detection in trajectory planning of robotic arms is used to determine whether the path point is valid, which is a key part of the trajectory planning algorithm. Planning a path in a working environment requires tens of thousands of collision detections, so the speed of collision detection has a great influence on the planning time. To accelerate collision detection, several learning-based proxy collision detectors are proposed. However, they do not perform well in accuracy, resulting in unsafe paths. In this paper, we model collision detection as a sequence estimation problem for the first time, and propose a novel Transformer-based high-accuracy collision distance estimator, DistFormer, which can be directly embedded in the trajectory planner for collision detection. DistFormer models robotic arm and obstacles as bounding sequences to retain shape information, and uses encoder blocks based on self-attention mechanisms to aggregate point information into edge information. The cross-attention mechanism is then used to fuse the feature sequences of the robotic arm and the obstacles. Finally, the self-attention mechanism is used to convert the robot sequence fused with position information of obstacles into the collision distance. Extensive experiments have shown that DistFormer achieves higher accuracy than all baselines in different working environments.