A Gap Inference Neural Network for Efficient Collision Detection in Robotic Arm Path Planning

With the widespread application of robots in industrial manufacturing and service sectors, the safety and efficiency of path planning in complex dynamic environments have become increasingly important. Collision detection is an essential step in path planning, yet traditional methods based on geometric models or physical simulations often face challenges such as low computational efficiency and inadequate adaptability to complex dynamic environments. In this paper, we propose a Gap Inference Network (GIN) for efficient collision detection in robotic arm path planning. This network aims to enhance the efficiency of collision detection during path planning in complex and dynamic environments. We employ a voxelization strategy to partition the robotic arm's workspace into uniformly distributed cubic units, and we collect an extensive dataset of gaps between the robotic arm and all cubic units corresponding to various joint configurations. The GIN is trained to learn the complex nonlinear mapping between the joint configurations and The Gap Field of all cubic units relative to the robotic arm. This enables rapid batch inference of The Gap Field for any given robotic arm joint configuration, significantly improving collision detection efficiency. Experimental results demonstrate that the GIN achieves an average collision prediction accuracy of 92.23 % and an average non-collision prediction accuracy of 98.44%, markedly enhancing the efficiency of collision detection.