Physics-guided Neural Network-based Friction Compensation Controller for Robotic Manipulators

This study proposes a physics-guided neural-network (PGNN) feedforward compensator for friction-related joint-torque residuals in robotic manipulators. The Franka Emika Panda robot’s low-pass-filtered external-torque estimate was used as the learning target during contact-free single-joint motion. Joint-wise multilayer perceptrons combine a soft direction penalty, tanh activations, and velocity history. Adding measured torque history reduced validation RMSE by 58–94% relative to the uncompensated residual, but generated outputs opposing the velocity direction for up to 48% of a steady-state input slice with the other inputs held fixed and an estimated offset removed. A configuration using joint position and five velocity taps up to 40 ms was selected for deployment at 1 kHz, reducing validation RMSE by 44–87%. Sine, step, and chirp tracking tests were repeated three times per condition. Over 1.5–20 s, the compensator reduced J6 and J7 tracking RMSE by 23–54%, compared with 13–44% for Gaz et al.’s sigmoid model. However, it increased error at J2 and J3 and increased J5 step-tracking RMSE by 28% under oscillation and saturation. Lower offline prediction errors alone did not consistently translate into better closed-loop tracking.