Multi-agent cooperative control for unmanned distributed-drive electric agricultural vehicle in paddy fields: tracking, stability, and energy-aware torque allocation

Abstract

With the rapid advancement of intelligent agriculture and autonomous field operations, the distributed drive electric plant protection vehicle (DDEPPV) is increasingly adopted for paddy-field plant protection. However, in soft-soil, low-adhesion, and highly disturbed environments, the tight coupling among path tracking, drive/yaw stability, and energy consumption-together with uncertain ground parameters-poses major challenges to conventional control. In addition, from-scratch reinforcement learning is difficult to deploy, as early exploration can induce yaw instability and wheel entrapment. To overcome these limitations, we propose a vehicle-level distributed electric-drive control framework that integrates physics-informed priors with multi-agent cooperative learning. A mud-water multiphase wheel-soil interaction model is built via CFD-DEM coupling to identify, under the parameter settings and operating conditions considered in this study, an energy- and sinkage-risk-aware slip-ratio window, thereby providing an interpretable ground-mechanics boundary for subsequent controller design. Under the centralized training and decentralized execution paradigm, the task is decomposed into three agents for path tracking, stability/traction regulation, and energy-optimal four-wheel allocation, and trained using model predictive control (MPC) expert-supervised pretraining followed by multi-agent twin delayed deep deterministic policy gradient (MATD3) cooperative fine-tuning. Real-time Hardware-in-the-Loop (HIL) experiments verify improved turning performance and enhanced yaw/traction stability, while reducing traction-system electrical energy consumption by 29.4% versus MPC and by an additional 5.4% over unpretrained MATD3, demonstrating unified optimization of accuracy-stability-energy efficiency in paddy fields.