Whole-Body Self-Collision Distance Detection for a Heavy-Duty Manipulator Using Neural Networks

Abstract

Many applications in manipulators require computing the minimum self-collision distance among links for safety. The calculation is time-consuming, especially when whole-body shapes are considered. To improve computational efficiency, a neural network-based hierarchical self-collision detection method is proposed, in which a classifier and a regressor are separately trained on binary collision labels and precise distance values respectively. The classifier swiftly filters out collision states, while the regressor focuses on predicting positive distances for collision-free states. Finally, geometric re-checking is triggered when the predicted distance falls below a tunable threshold. We evaluate the accuracy, computational efficiency, and safety of our approach through extensive experiments on a manipulator of tunneling drilling rig. The results demonstrate that at a 4cm threshold, our technique achieves 6.23% of the computational cost of state-of-the-art geometric checkers while maintaining high safety. Adjusting the threshold allows for a trade-off between efficiency and safety.