Compared to conventional fuel-powered agricultural machinery, the "new energy + electric drive" system demonstrates significant benefits in energy efficiency and emission management. Nevertheless, field operations involve large load fluctuations and complex soil conditions, presenting multiple challenges for electric-driven agricultural equipment, including energy management, drive system matching, and reliability control during prolonged continuous operation. Accordingly, we developed a distributed electric-drive unmanned agricultural robot (TerraVolt) and proposed an energy-efficient control strategy enabled by digital twin (DT). Firstly, considering terrain information, we developed the dynamic mechanism model of the TerraVolt driving system, named Virtual-TerraVolt. Secondly, we designed a wheel longitudinal force estimation algorithm based on a single-wheel longitudinal dynamics model and a hyper-spiral observer, and proposed a particle filter (PF) based on the wheel-soil model to estimate terrain parameters. Then, an improved deep deterministic policy gradient algorithm with expert-guided soft-hard constraint module (DDPG-SH) was developed to achieve intelligent torque allocation of the drive motors. Finally, the strategy was validated through simulation experiments and field tests. Simulation results indicate that the proposed strategy decreases energy consumption by 7.19% relative to rule-based speed controllers. Compared to the traditional DDPG algorithm, speed stability was enhanced by 36.77%. Field results show that the proposed strategy improved speed stability by 43.21% compared to the classic DDPG algorithm, and reduced energy consumption by 10.26% compared to the operation speed controller. This study offers an effective and intelligent solution for the high-efficiency operation of smart agricultural machinery and promotes the deep integration of DT and reinforcement learning (RL) technologies in agricultural machinery control domain.
Wang et al. (Mon,) studied this question.