Title: CropCraft: A Procedural World Generator for Robotic Simulation of Agricultural Tasks

Abstract:

As modern agriculture increasingly embraces agroecological practices, there is a growing need for robotic systems capable of navigating and operating within highly complex and diverse field conditions. While simulation is essential for developing and testing these systems, creating realistic, configurable 3D environments that accurately reflect agroecological variety remains a significant hurdle. To address this, we introduce CropCraft, an open-source procedural world generator developed using Python and Blender. This tool is specifically designed to create 3D simulation environments customized for agricultural robotics.

By utilizing a straightforward YAML configuration file, CropCraft generates crop fields suitable for a broad spectrum of scenarios, such as vineyards, fields affected by weeds, and intercropping setups. The system features a comprehensive library of 3D plant models—including crops, grasses, and weeds—depicted at various stages of growth. To mimic the spatial variability found in actual agricultural fields, the generator employs stochastic placement algorithms. Furthermore, the resulting worlds can be directly imported into the Gazebo simulator and come equipped with ground-truth annotations for every placed element, facilitating the development of both perception and navigation algorithms.

To illustrate the practical value of CropCraft, we applied it to the challenge of crop-weed semantic segmentation using deep learning techniques. We generated a dataset comprising 10,000 synthetic images of maize fields, incorporating variations in lighting conditions, weed densities, and plant growth stages. Several segmentation architectures were trained on this synthetic data. Notably, models trained solely on these synthetic images achieved a sim-to-real gap of roughly 10% mean Intersection over Union (mIoU) when tested on real field images, surpassing previous state-of-the-art synthetic generation methods. Additionally, our results indicate that integrating even a small number of real images with synthetic data enhances domain generalization, offering fresh perspectives on the optimal utilization of synthetic data for agricultural perception tasks.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC