GreenhouseSplat: A Dataset of Photorealistic Greenhouse Simulations for Mobile Robotics

TL;DR

The paper tackles the lack of photorealistic, greenhouse-scale simulation for mobile robotics by introducing GreenhouseSplat, a pipeline that generates photorealistic greenhouse assets from inexpensive RGB images using 2D Gaussian Splatting and integrates them into a ROS-based simulator. It demonstrates a complete workflow from real-image capture to Gaussian-based reconstruction, rendering, and sensor simulation (camera and LiDAR), coupled with a proof-of-concept localization task using fiducial markers. The authors provide a dataset of 82 cucumber plants across eight row-end segments to enable robotics evaluation in varied configurations, establishing a foundation for greenhouse-scale radiance-field simulation and sim-to-real research in agricultural robotics. Overall, GreenhouseSplat offers a practical path to realism-driven evaluation, lowers the barrier to greenhouse-specific perceptual testing, and highlights directions for scaling to broader crops and environments.

Abstract

Simulating greenhouse environments is critical for developing and evaluating robotic systems for agriculture, yet existing approaches rely on simplistic or synthetic assets that limit simulation-to-real transfer. Recent advances in radiance field methods, such as Gaussian splatting, enable photorealistic reconstruction but have so far been restricted to individual plants or controlled laboratory conditions. In this work, we introduce GreenhouseSplat, a framework and dataset for generating photorealistic greenhouse assets directly from inexpensive RGB images. The resulting assets are integrated into a ROS-based simulation with support for camera and LiDAR rendering, enabling tasks such as localization with fiducial markers. We provide a dataset of 82 cucumber plants across multiple row configurations and demonstrate its utility for robotics evaluation. GreenhouseSplat represents the first step toward greenhouse-scale radiance-field simulation and offers a foundation for future research in agricultural robotics.

Figures (5)

• Figure 1: Overview of the GreenhouseSplat pipeline. Input RGB images are processed with a pre-trained MAST3R model to obtain feature matchings, followed by Structure-from-Motion (SfM) with manual alignment to recover camera poses and a sparse point cloud. These are used to train a 2D Gaussian Splatting model, producing Gaussian primitives. After post-processing and cleanup, we obtain photorealistic GreenhouseSplat assets suitable for simulation.
• Figure 2: Greenhouse environment used for data collection
• Figure 3: Sparse reconstruction with MAST3R-SfM of one row end from the greenhouse. The reconstruction is aligned with the $z$-axis pointing upwards and the $x$-axis oriented along the row end.
• Figure 4: LiDAR simulated by back-projection of depth maps from gaussian splatting. Point cloud along with robot model visualized in RViz
• Figure 5: Proof-of-concept localization task in the simulated greenhouse environment.