A virtual tomato training arena for harvesting robots

Researchers at Osaka Metropolitan University have created a system that generates realistic virtual tomato farms to aid in training artificial intelligence for agricultural applications.
The virtual farm environment uses images from real-world farms obtained through agricultural robots to accurately replicate real-world conditions.
Advanced reconstruction techniques and 3D Gaussian Splatting were employed to simulate complex scenarios like overlapping leaves, dynamic lighting, and hidden fruits.
The AI model automatically generates labels for ripeness and annotates objects in YOLO format, easing the laborious process of manually tagging data for training.
The team demonstrated that AI models trained on their synthetic data could accurately detect tomatoes in real farm images.
The study provides insights into how factors like lighting, tomato shape, and dataset size impact AI accuracy.
The researchers believe this method can potentially be applied to harvesting a variety of crops beyond just tomatoes.
The findings were published in the journal Smart Agricultural Technology.