State-space feature engineering for GPR-based feedforward compensation in a 5-DOF robotic manipulator

Analytical rigid-body models of robotic manipulators neglect friction and load-dependent disturbances, and the residual torque that remains degrades tracking accuracy. This paper compares Gaussian process regression (GPR) and artificial neural networks (ANN) as estimators of that residual on a five degree-of-freedom Mitsubishi Movemaster RV-M1. The residual is the torque inferred from armature current minus the analytical Euler–Lagrange model. The main finding is that generalization is governed by the input representation rather than by the model class. Indexed by elapsed time, both learners fail beyond the training horizon, with test errors of 4.1 to 11.2 N m. Re-indexing the same GPR by the kinematic state reduces the test error by 55.7% at Joint 2 and by more than 99% at the remaining joints over five rolling origins. A two-by-two ablation identifies the representation, not the kernel, as the dominant factor. The accuracy obtained at four joints follows from the residual being a static function of joint velocity to within 0.1% of its variance. Joint 2 lacks this structure: its predictive interval is miscalibrated, and a searched network matches GPR. Trained on one direction of travel and tested on the other, both representations perform worse than a mean predictor. In a closed-loop analysis driven by the measured residual, the estimator used as a feedforward compensator removes 72% to 99% of the PID feedback effort at Joints 1 to 4 over a 0 to 1.5 kg payload range.