Inverse Kinematics Analysis of Baxter Robot Based on Residual BP Neural Network

A self-adaptive inverse kinematics analysis method based on the residual back propagation (BP) neural network was proposed. Around the entire motion planning process from data acquisition to real-time control, 140 sets of position and Euler angle data were collected. The residual BP neural network was employed to perform inverse kinematics analysis on Baxter robot’s arm, and 7 joint angles of the robot’s arm were fitted. Additionally, the trained joint angles were published in the form of topics, and realized the communication by subscribing to the topic in the script for grasping objects. Combined with visualization with Rviz and real-world dual-arm cooperative experiments, grasping experiments were conducted on four object models with residual BP neural network and ordinary BP neural network, respectively, to verify the effectiveness of the proposed method. The results show that the calculation time for a single point of the proposed method is approximately 8.1 ms, which is much shorter than the control cycle of the robotic arm, and can achieve real-time requirements. With 1 500 rounds of training, the mean square error of the residual BP neural network model is 0.006, which is 0.077 lower than that of the ordinary BP neural network model, thus improves the accurary of the model. The success rate of the proposed method for grasping is 87.5%, which is 22.5% higher than that of ordinary BP neural network, verifying the effectiveness and practicality of the method proposed in this paper.