Learning-Based Adaptive Robust Control of a 4-DOF Robot Manipulator

This paper presents a learning-based adaptive control strategy for trajectory tracking of a four-degree-of-freedom (4-DOF) robot manipulator subject to unknown disturbances and model uncertainties. The designed controller combines nonlinear model-based compensation with online adaptation of the feedback gains and of a lumped disturbance estimate, both driven by the deviation of the actual tracking-error dynamics from the prescribed target error dynamics. A local stability condition is established for the resulting error dynamics. In simulations of the 4-DOF Quanser QArm, the controller outperformed conventional feedback linearization and an equal-bandwidth proportional–integral–derivative (PID) controller with gravity compensation on all joints, including under ±20% parametric uncertainty, payload variation, and measurement noise, with most of the improvement resulting from the disturbance estimate. Hardware experiments, limited to the wrist joint of the QArm, demonstrated that the controller can be executed in real time through a current-command interface with bounded adaptive variables and smoother currents than the vendor-supplied PID controller. Across the complete experiments, the designed controller achieved lower tracking error in four of the five tested conditions, demonstrating improved tracking performance across most of the evaluated scenarios. An additional analysis excluding the start-up period of the vendor PID controller showed that the PID achieved 11–51% lower tracking error in three unloaded tests, whereas the designed controller achieved 27–36% lower error in two payload tests conducted with modified settings. These additional results highlight the influence of the evaluation period and operating conditions on the relative performance of the two controllers, while the overall comparison favors the designed controller in the majority of the tested conditions. A wrist model identified from the experimental data shows that static friction at very low currents, together with the evaluation of the learning signal without acceleration measurement, explains most of the difference between the simulation and experimental results.