Deep reinforcement learning–driven model-based adaptive PID strategy for six-wheeled rocker–bogie mobile platform

Sažetak

Introduction Six-wheel rocker–bogie rovers encounter nonlinear dynamics, variable traction, and wheel slip that can degrade the performance of fixed-gain proportional–integral–derivative (PID) controllers. This study develops a hybrid deep reinforcement learning–PID framework to improve steering and drive control under these conditions. Methods Two Deep Deterministic Policy Gradient (DDPG) actor–critic agents adapt the steering and drive PID gains online using low-level dynamic feedback. The reward functions penalise tracking errors, wheel slip, inter-joint mismatch, and abrupt control variations. The framework was evaluated in a three-dimensional MATLAB Simscape Multibody environment incorporating suspension articulation, wheel–terrain contact, and load redistribution. Comparisons with conventional PID and standard DDPG-PID controllers examined tracking accuracy, learning convergence, actuator responses, and disturbance rejection. Results The proposed framework demonstrated more consistent learning convergence, reduced tracking deviations, and smoother actuator torque responses than the baseline controllers. Online gain adaptation improved the steering and drive responses under simulated variations in traction, slip, and payload distribution. The shaped rewards reduced fluctuations in gain updates and critic estimates. Discussion These simulation findings support the potential of reinforcement learning–based PID gain adaptation for low-level control of rocker–bogie rovers under uncertain operating conditions. The framework retains the conventional PID structure while introducing data-driven adaptation. Hardware-in-the-loop testing and physical rover experiments remain necessary to establish performance beyond simulation.