Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluating Path Planning Strategies for Efficient Nitrate Sampling in Crop Rows

Ruiji Liu, Abigail Breitfeld, Srinivasan Vijayarangan, George Kantor, Francisco Yandun

TL;DR

Problem addressed: efficient nitrate sampling in crop rows requires fast, reliable local path planning under crop-row constraints. Approach: compare A* graph search, DQN-based reinforcement learning, and a heuristic method within a UAS-derived orthomosaic global map and GPS-based ground navigation on the Amiga robot. Key findings: heuristic planning achieves 0.28 ms and 100% success, A* is near-optimal, and DQN shows longer delays with occasional suboptimal routes. Significance: deterministic, geometry-aware planners excel in structured agricultural environments and motivate future hybrid and automation efforts for precision agriculture.

Abstract

This paper presents a pipeline that combines high-resolution orthomosaic maps generated from UAS imagery with GPS-based global navigation to guide a skid-steered ground robot. We evaluated three path planning strategies: A* Graph search, Deep Q-learning (DQN) model, and Heuristic search, benchmarking them on planning time and success rate in realistic simulation environments. Experimental results reveal that the Heuristic search achieves the fastest planning times (0.28 ms) and a 100% success rate, while the A* approach delivers near-optimal performance, and the DQN model, despite its adaptability, incurs longer planning delays and occasional suboptimal routing. These results highlight the advantages of deterministic rule-based methods in geometrically constrained crop-row environments and lay the groundwork for future hybrid strategies in precision agriculture.

Evaluating Path Planning Strategies for Efficient Nitrate Sampling in Crop Rows

TL;DR

Problem addressed: efficient nitrate sampling in crop rows requires fast, reliable local path planning under crop-row constraints. Approach: compare A* graph search, DQN-based reinforcement learning, and a heuristic method within a UAS-derived orthomosaic global map and GPS-based ground navigation on the Amiga robot. Key findings: heuristic planning achieves 0.28 ms and 100% success, A* is near-optimal, and DQN shows longer delays with occasional suboptimal routes. Significance: deterministic, geometry-aware planners excel in structured agricultural environments and motivate future hybrid and automation efforts for precision agriculture.

Abstract

This paper presents a pipeline that combines high-resolution orthomosaic maps generated from UAS imagery with GPS-based global navigation to guide a skid-steered ground robot. We evaluated three path planning strategies: A* Graph search, Deep Q-learning (DQN) model, and Heuristic search, benchmarking them on planning time and success rate in realistic simulation environments. Experimental results reveal that the Heuristic search achieves the fastest planning times (0.28 ms) and a 100% success rate, while the A* approach delivers near-optimal performance, and the DQN model, despite its adaptability, incurs longer planning delays and occasional suboptimal routing. These results highlight the advantages of deterministic rule-based methods in geometrically constrained crop-row environments and lay the groundwork for future hybrid strategies in precision agriculture.

Paper Structure

This paper contains 15 sections, 1 equation, 6 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Global mapping and navigation
  4. Aerial Survey
  5. Ground Robot Controller
  6. Planning Problem
  7. Graph Search Formulation
  8. Reinforcement Learning Formulation
  9. DQN Model
  10. Heuristic Search
  11. Experimental Results
  12. Experiments Setup
  13. Simulation Deployment
  14. Discussion
  15. Conclusion

Figures (6)

  • Figure 1: Robot designed for sampling nitrate content in corn fields. It features a robotic arm equipped with a custom gripper for nitrate probing, navigation sensors, and an onboard computer.
  • Figure 2: Global navigation layout. A pre-computed map of the field is used to define the ground robot's path. Then it tracks it to reach designated nitrate sampling points.
  • Figure 3: Overview of A* Graph search: (a) possible actions based on the robot’s position in the field, and (b) an example planned trajectory.
  • Figure 4: Custom Gym environment (CropRowEnv) for crop row path planning RL agent training (10 crop rows x 10 units length), with a two-component action space.
  • Figure 5: Example deployment of output actions [0,0], [1,7], [1,0] in a simulated field. The blue star represents the goal sampling point.
  • ...and 1 more figures