Temporal compensation physics-informed neural network modeling repetitive small-range rotation inverse dynamics.

High-fidelity motion mapping is fundamental for high-precision robotic operations. However, existing models exhibit dynamic hysteresis effects during repetitive small-range rotations, resulting in diminished modeling accuracy. Therefore, we propose a temporal compensation physics-informed neural network (TC-PINN) to mitigate the dynamic hysteresis effects. First, we construct a rotational state trajectory encoded with cumulative joint rotation features to explicitly characterize the temporal phase deviations of motion hysteresis. Second, we design a Linformer-based nonlinear error fitting structure to compensate for unstructured dynamic errors caused by dynamic hysteresis effects. Finally, we establish TC-PINN deployed for end-to-end inference, which is optimized through a staged training strategy to mitigate mapping misalignment. Experimental results on a seven degrees of freedom robotic arm show that TC-PINN improves hysteresis-aware torque mapping accuracy under repetitive small-range rotations, reducing the total mean absolute error to 9.65×10-4 Nm.