Data-Driven Identification of Robot DH Parameter Errors Using a Physics-Informed Transformer Network

The pose accuracy of a six-axis robotic arm is crucial for high-end applications such as aerospace and precision manufacturing. For robotic arms using DH (Denavit-Hartenberg) for kinematic modeling, model parameter errors originating from machining tolerances are the key factors limiting their positioning accuracy. Traditional parameter calibration methods using iterative optimization are often sensitive to initial values and prone to converging to local optimal solutions. This study proposes a neural network-based parameter identification method. Using pose error (represented by axis-angle notation) and joint angles as inputs, the parameter alpha is identified as output and undergoes standardization. The transformer encoder serves as the network structure, with a physical loss term derived from forward kinematics calculations added to the loss function. The final trained model achieved a Mean Absolute Error (MAE) of 2.5×10-4 (rad) in predicting alpha on 50 test samples, a 37.78 % reduction compared to uncompensated values. Using the compensated parameters, the pose error was reduced by 94.57 % compared to the uncompensated kinematic model, demonstrating the effectiveness of the proposed method.