Local Navigation and Docking of an Autonomous Robot Mower Using Reinforcement Learning and Computer Vision

Ali Taghibakhshi, Nathan Ogden, Matthew West

Introduction

Reliable docking represents a critical "last-mile" challenge for autonomous service robots, particularly for outdoor systems like the John Deere Tango mower. Traditional commercial solutions often rely on buried boundary wires that guide the mower via induction. However, these systems are prone to failure if the wire shifts, the ground is uneven, or obstacles block the path. While high-end sensors like LIDAR or RTK-GPS offer better precision, they are often prohibitively expensive for consumer-grade hardware.

This research presents a low-cost, vision-only system designed to navigate a mower to its charging station from arbitrary starting positions and orientations. By combining deep learning-based object detection with reinforcement learning (RL), the system maintains high accuracy even in variable environmental conditions without the need for expensive specialized sensors.

The physical experimental setup, showing the John Deere mower with markers, the charging station, and the stationary camera.

Methodology

The proposed system utilizes a two-stage pipeline that bridges the gap between raw visual data and robotic control. This approach focuses on dimensionality reduction to create a robust "sim-to-real" bridge.

Feature Extraction via YOLO

Instead of feeding high-dimensional raw pixels directly into a Reinforcement Learning agent, the system employs a YOLO (You Only Look Once) network. This network is trained to detect two specific "opportunistic" markers already present on the mower: the "stop" button and the John Deere logo. By tracking these two points, the system distills complex visual data into four simple coordinates ($x, y$ for each marker), effectively determining both the mower's position and heading relative to the docking station camera.

Control Logic and Training

The coordinates from the YOLO network—spanning the last three time-steps to capture temporal context like velocity and acceleration—are fed into a Double Deep Q-Network (Double DQN) agent. The use of a Double DQN helps prevent the overestimation of action values, leading to more stable learning.

The neural network architecture, illustrating the flow from visual input to marker detection and finally to action selection via the RL agent.

The agent was trained using a custom-built simulation environment where mower kinematics were modeled as Ordinary Differential Equations (ODEs). A four-phase curriculum training strategy was employed, starting with simple rewards for forward movement and progressing to stringent requirements for precise docking coordinates and minimal angular offset.

Experimental Results

The system was validated through 90 real-world hardware tests covering various initial positions and orientations. The results demonstrated centimeter-level precision:

  • Position Accuracy: The mean absolute error was less than 1 cm in both the X and Y directions, with a maximum recorded error under 4 cm.
  • Orientation Accuracy: The mean orientation error was 1.53°, with a maximum error of 6.2°.
  • Robustness: The mower successfully docked from arbitrary initial angles (up to ±30°) and offsets, proving significantly more flexible than traditional wire-following methods.

Discussion and Future Work

One of the primary insights of this work is that dimensionality reduction serves as an effective sim-to-real bridge. By reducing the state space to marker coordinates, the transition from simulation to physical hardware was remarkably smooth. Furthermore, leveraging existing design elements as markers demonstrates that vision-only systems can be integrated into existing products with minimal physical modification.

Future research will explore mounting the camera directly on the mower to navigate toward a stationary target on the dock, which would remove the need for a fixed external camera. Additionally, the use of custom-designed, high-contrast markers and testing in more diverse weather conditions could further improve the system's robustness for commercial deployment.

Summary updated

Abstract

We demonstrate a successful navigation and docking control system for the John Deere Tango autonomous mower, using only a single camera as the input. This vision-only system is of interest because it is inexpensive, simple for production, and requires no external sensing. This is in contrast to existing systems that rely on integrated position sensors and global positioning system (GPS) technologies. To produce our system we combined a state-of-the-art object detection architecture, You Look Only Once (YOLO), with a reinforcement learning (RL) architecture, Double Deep Q-Networks (Double DQN). The object detection network identifies features on the mower and passes its output to the RL network, providing it with a low-dimensional representation that enables rapid and robust training. Finally, the RL network learns how to navigate the machine to the desired spot in a custom simulation environment. When tested on mower hardware, the system is able to dock with centimeter-level accuracy from arbitrary initial locations and orientations.
BibTeX
@inproceedings{taghibakhshi2021local,
  author = {Ali Taghibakhshi and Nathan Ogden and Matthew West},
  title = {{Local Navigation and Docking of an Autonomous Robot Mower Using Reinforcement Learning and Computer Vision}},
  booktitle = {ICCAE},
  year = {2021},
  url = {https://openreview.net/forum?id=Sb8dri34pU5},
}
Figure