Back propagation neural network based real-time self-collision detection method for humanoid robot

This paper proposes a back propagation neural network based real-time humanoid self-collision detection method which eliminates the repetition of detection computation for same and similar motion sets. The proposed system is able to reduce self-collision detection computation time significantly, because of the pattern recognition capability of the neural network. However, the accuracy of back propagation neural network based self-collision detection cannot be guaranteed 100%. For this reason, the system is also designed to detect potential miss detected motion sets though the module based self-collision detection method, which eliminates unnecessary motion pairs by focusing on certain modules with higher collision probability. Our module based self-collision detection method is a failsafe method. The proposed method has been implemented on a humanoid simulator, which modeled selected humanoid motion. The performance of our method successfully reduces the computation time of self-collision detection, and the “Fail safe” also operates successfully. For this reason, our method can improve the real-time motion control of humanoids in uncertain environment.