Reinforcement Learning-Driven Robotic Swarm Coordination for Precision Agriculture: A Longitudinal Field Study on Yield Optimisation and Operational Autonomy

Abstract

Distributed IoT sensing and multi-agent reinforcement learning (MARL) present largely unexplored synergies for precision agriculture automation, yet no peer-reviewed work has demonstrated their integration in a longitudinal field deployment with heterogeneous robotic fleets. This article reports the design and 18-month field evaluation of a MARL-driven robotic swarm coordination system deployed across a 340-hectare multi-crop research farm. The system integrates a 20-robot heterogeneous fleet—12 ground robots and 8 aerial drones—with an 847-node distributed IoT sensor network and a composite agronomic reward formulation for MARL policy training under a centralized training with decentralized execution (CTDE) paradigm. Over 18 months spanning two complete growing seasons, the system achieved a 23.4% mean crop yield improvement (4.2 to 5.2 t/ha; t(3) = 6.17, p = 0.009, Cohen's d = 3.08) and a 67.0% reduction in irrigation water consumption (847 to 280 m³/ha/season; 95% CI: [62.4%, 71.2%]), alongside a 15-fold improvement in pest stress detection lead time. Soil moisture was maintained within ±3.2% of the crop-optimal range 91.4% of the time, producing a water use efficiency of 2.04 kg/m³—a 2.84× improvement over the 0.72 kg/m³ pre-automation baseline. An event-driven adaptive sampling protocol extended median sensor battery life 3.4-fold (from 4.2 to 14.3 months). Swarm coordination efficiency reached 94.7% of urgent tasks completed within the 6-hour agronomic response window, with inter-agent conflict rates declining from 3.2 to 0.8 per 100 operating hours across the first six months through online MARL policy adaptation. Deployment break-even economics are confirmed at month 31 under conservative NPV assumptions. These results establish MARL-orchestrated heterogeneous robotic fleets as a technically and operationally validated approach to precision agriculture automation.