Procedural generation of synthetic forest environments to train machine learning algorithms

The demand for the development of forestry robotics has been increasing. As with most robotics applications, Machine Learning is the engine driving innovation in this field. However, Machine Learning development for robotic perception tasks is highly dependent on the availability of annotated datasets. Contrasting with urban environments, public datasets for forest applications are rare, hard to collect and currently not enough to train models capable of operating autonomously. This paper proposes a solution to mitigate the data shortage problem: a system that uses procedural generation to create virtual forests and collects synthetic data from these environments using virtual sensors. More specifically, the system generates RGB images and point clouds with pixel-wise and point-wise annotations, respectively, as well as depth maps, substantially reducing the time and effort invested in dataset construction. The system proved capable of generating 1000 frames with all the above-mentioned data types in 3 hours of autonomous operation. The generated data is ready to be used in Machine Learning model training. Finally, qualitative preliminary results obtained by a semantic segmentation model trained on the generated dataset, which has been made publicly available in a community-wide repository, are presented