Reinforcement Learning-Guided Model Predictive Control for Economic Velocity Optimization of Hybrid Electric Tracked Vehicles Considering Steering Characteristics

Unlike wheeled vehicles' eco-driving that focuses on longitudinal dynamics, tracked vehicles adopt skid-steering with significant steering resistance moment, making the power demand dependent on both longitudinal and steering dynamics. However, accurately acquiring the steering resistance moment and leveraging it to coordinate fuel economy and vehicle mobility remains a critical challenge. To address this problem, a reinforcement learning (RL)-guided predictive eco-driving strategy is proposed, consisting of three modules: a neural network-based steering characteristics model, a time reference generator, and an RL-guided rolling optimization module. First, a neural network-based model is trained using data from the dynamics model to learn the relationship between steering resistance moment and vehicle motion states. Second, a time reference generator is designed to enable the vehicle to reach the destination on time. Finally, the eco-driving problem is formulated in the MPC framework and solved via RL-guided hierarchical particle swarm optimization (RLHPSO), where RL adaptively tunes HPSO levels for superior solutions. Simulation results show the proposed strategy improves fuel economy by 6.99% and mobility by 3.92% over the strategy neglecting steering characteristics, and outperforms MPC-HPSO by 4.99% in economy and 2.07% in mobility. An additional off-road scenario further confirms the generalizability of the proposed strategy. Hardware-in-the-loop tests demonstrate its promising online applicability.