Reinforcement Learning-Based Energy-Aware Coverage Path Planning for Precision Agriculture

TL;DR

The paper tackles energy-constrained coverage path planning (CPP) for autonomous agricultural robots operating in grid-based fields with obstacles and charging stations. It introduces a Soft Actor-Critic (SAC) framework that uses a CNN-LSTM architecture to extract spatial and temporal features and optimizes a multi-objective reward balancing coverage, energy usage, and safe return. The environment state is represented as a 4-channel tensor $\mathcal{S}_t \in \mathbb{R}^{4 \times N \times N}$ with four cardinal actions, enabling end-to-end learning of energy-aware CPP policies. Experiments across multiple map configurations demonstrate >90% coverage and substantially fewer energy-constraint violations than baseline heuristics (RRT, PSO, ACO), highlighting the method’s robustness and practical potential for precision agriculture. The work presents a scalable, energy-aware CPP solution that integrates learning-based decision making with safety guarantees to support autonomous agricultural robotics in real-world deployments.

Abstract

Coverage Path Planning (CPP) is a fundamental capability for agricultural robots; however, existing solutions often overlook energy constraints, resulting in incomplete operations in large-scale or resource-limited environments. This paper proposes an energy-aware CPP framework grounded in Soft Actor-Critic (SAC) reinforcement learning, designed for grid-based environments with obstacles and charging stations. To enable robust and adaptive decision-making under energy limitations, the framework integrates Convolutional Neural Networks (CNNs) for spatial feature extraction and Long Short-Term Memory (LSTM) networks for temporal dynamics. A dedicated reward function is designed to jointly optimize coverage efficiency, energy consumption, and return-to-base constraints. Experimental results demonstrate that the proposed approach consistently achieves over 90% coverage while ensuring energy safety, outperforming traditional heuristic algorithms such as Rapidly-exploring Random Tree (RRT), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO) baselines by 13.4-19.5% in coverage and reducing constraint violations by 59.9-88.3%. These findings validate the proposed SAC-based framework as an effective and scalable solution for energy-constrained CPP in agricultural robotics.

Figures (4)

• Figure 1: Attention-augmented dual-convolutional architecture for energy-constrained CPP.
• Figure 2: Three different maps used in our experiments. Green cells indicate the charging stations and the black cells indicate the obstacles. Each cell represents a 1m $\times$ 1m area in the agricultural field. Map 1 (left) shows a simple environment with minimal obstacles and evenly distributed charging stations. Map 2 (center) presents a more complex layout with clustered obstacles representing buildings and equipment. Map 3 (right) features a challenging environment with irregular obstacle patterns and strategic charging station placement.
• Figure 3: Rewards in three training maps. The x-axis represents the number of training episodes, and the y-axis shows the cumulative reward values. Higher reward values indicate better performance in balancing coverage efficiency and energy consumption. Map 1 shows faster convergence due to simpler terrain, while Maps 2 and 3 demonstrate more gradual improvement as the agent learns to navigate more complex environments.
• Figure 4: Percentages of coverage in three training maps. The x-axis represents the number of training episodes, and the y-axis shows the percentage of accessible areas successfully covered by the robot. Map 1 achieves near-complete coverage (>95%) after 400 episodes. Maps 2 and 3 show more variability but consistently reach above 90% coverage in later training stages despite increased environmental complexity.