AI for Green Spaces: Leveraging Autonomous Navigation and Computer Vision for Park Litter Removal

Christopher Y. Kao, Akhil Pathapati, James Davis

TL;DR

This paper presents a novel robotic system designed to autonomously navigate, identify, and collect litter in park environments. By combining a Spanning Tree Coverage path planning algorithm with RTK-GPS for precise localization and a ResNet50 CNN for trash detection, the robot achieves an impressive 80% overall success rate in clearing grass fields.

What Problem Does This Paper Address?

Litter in public spaces, particularly grass fields, poses a significant environmental and aesthetic challenge. While robotic solutions exist for beaches and water, the unique challenges of grass terrain—such as the need to avoid damaging living vegetation and the dispersed nature of litter—have limited autonomous solutions. Existing research often relies on simulations, human-robot collaboration, or robots with limited trash detection capabilities, leaving a gap for a fully autonomous, robust system capable of operating effectively in natural, uneven environments.

Methods

The authors developed a multi-faceted robotic system with three core components:

  • Autonomous Navigation:
    • Path Planning: They employed the Spanning Tree Coverage (STC) algorithm. Unlike Simultaneous Localization and Mapping (SLAM), STC generates an efficient, non-overlapping coverage path by treating the area as a grid and creating a spanning tree, ensuring the entire area is traversed only once.
  • Localization: To achieve centimeter-level accuracy required for precise navigation, they utilized Real-Time Kinematic (RTK) GPS. This system corrects GPS signal errors by using a nearby base station, providing highly accurate positioning data. To further enhance accuracy and compensate for GPS signal latency, they incorporated a gyroscope and dead reckoning, along with a novel yaw drift correction algorithm that uses sequential GPS readings.
  • Computer Vision Trash Detection:
    • They opted for an image classification approach using a ResNet50 Convolutional Neural Network (CNN), defining two classes: "trash" and "no-trash." This broad classification allows the model to generalize to a wider variety of trash shapes and sizes compared to image segmentation models that are limited to predefined classes.
  • To overcome the limitations of existing datasets (like TACO) which have limited trash classes and often use artificial backgrounds, they created a custom dataset of 18,154 images of trash on grass, augmented with various techniques to simulate diverse environmental conditions.
  • Trash Pickup Mechanism:
    • After testing several designs, they selected a bow rake-inspired mechanism. This mechanism effectively sweeps litter resting on top of the grass without disturbing the vegetation. To address the issue of transporting collected trash, they integrated a brush to sweep the collected litter into an onboard receptacle.

Datasets

  • Custom Trash Dataset: A new dataset of 18,154 images was created, specifically featuring trash on grass backgrounds. This dataset was augmented to account for variations in grass color, texture, and lighting conditions.
  • TACO Dataset: Mentioned as an existing dataset, but noted as insufficient for their needs due to its limited classes and artificial backgrounds.

Key Insights and Ideas

  • STC for Efficient Coverage: The choice of STC over SLAM for path planning is a key insight, demonstrating that for known areas, a deterministic coverage algorithm can be more efficient by avoiding redundant traversals.
  • RTK-GPS for Precision: The successful implementation of RTK-GPS highlights its critical role in enabling precise outdoor navigation for robots, overcoming the limitations of standard GPS.
  • Broad Classification for Generalization: Using a simple "trash" vs. "no-trash" classification with ResNet50 proved more effective for detecting a wide variety of litter than more granular segmentation approaches, especially in a natural environment.
  • Rake Mechanism's Simplicity and Effectiveness: The bow rake mechanism's success demonstrates that a simple, inspired design can outperform complex robotic grippers for specific tasks like picking up light litter on grass.

Key Results

  • Overall Success Rate: The robot achieved an 80% overall success rate in clearing trash from the test area.
  • Autonomous Navigation: Achieved 100% success rate in covering the designated area without errors.
  • Trash Detection: Achieved a 90% success rate in detecting trash when driven over.
  • Trash Pickup: Achieved an 89% success rate in collecting detected trash.
  • ResNet50 Performance: The ResNet50 model achieved a 94.52% test accuracy for trash detection, outperforming MobileNet and EfficientNet on their custom dataset.
  • Navigation Accuracy: The yaw drift correction algorithm significantly improved the robot's path accuracy, reducing location estimation errors.

Limitations and Future Work

  • Trash Pickup Limitations: The robot struggled with larger paper bags, suggesting that the pickup mechanism may need to be enlarged. The detection of very small items like bottle caps was also noted as a potential issue, though deemed less critical due to their infrequent occurrence without bottles.
  • False Positives in Vision: The computer vision model produced false positives due to lighting changes and shadows. This could be mitigated by increasing the confidence threshold for trash detection or by training the model to ignore shadow variations.
  • Future Work: The authors suggest enlarging the pickup mechanism and improving the trash detection model's robustness to lighting variations. They also believe this pickup method could be a viable solution for current park maintenance challenges.

Summary updated

BibTeX
@inproceedings{kao2025ai,
  author = {Christopher Y. Kao and Akhil Pathapati and James Davis},
  title = {{AI for Green Spaces: Leveraging Autonomous Navigation and Computer Vision for Park Litter Removal}},
  booktitle = {Sii},
  year = {2025},
  doi = {10.1109/sii59315.2025.10870583},
}
Figure