SynthField: A Multimodal Synthetic Dataset for Agricultural Robotics in Row Crops Fields

The development of autonomous agricultural robots is often hindered by the seasonal availability of field data and the labor-intensive nature of semantic labeling. To address this, we present SynthField, a high-fidelity multimodal synthetic dataset generated in NVIDIA Isaac Sim (v6) and recorded via ROS 2 bags.

This dataset provides perfectly annotated, 4D phenotyping data alongside realistic navigation challenges to accelerate the training and testing of robust perception and navigation algorithms in agriculture.

---

🌾 Dataset Composition & Asset Library

The scenes are assembled using a custom procedural generation script that constructs the field according to realistic row and plant spacing metrics for each specific crop type, followed by the stochastic placement of weeds and obstacles.

Plant crop assets are sourced from the Xfrog Agriculture Library and were manually edited in Blender to physically separate branch meshes from leaf meshes for traversability annotation. Other obstacle and environmental models were collected from open-source repositories.

The dataset features four major crop types at different growth stages, along with various weeds and obstacles:

• Crops (9 distinct growth stages): Soybean, Corn, Cotton, and Sorghum.
• Weeds: Broadleaf (2 variations) and Grass (2 variations).
• Obstacles: Humans (2 variations), Tractors, Traffic Cones (4 variations), and a Wild Boar (serving as a true out-of-distribution novelty obstacle).
• Environmental Variability: Simulated sunrise-to-sunset lighting cycles and 3 distinct, high-resolution soil textures applied stochastically.

---

🚜 Acquisition Platform & Sensors

Data is captured using a simulated 4-wheel steering (4WS) acquisition platform.

The platform is equipped with a sensor suite designed to mimic standard autonomous field setups:

• RGB-D Camera: A front-facing camera modeled after the Stereolabs ZED-X, outputting synchronized RGB images, depth maps, and semantic masks.
• LiDAR: A 32-beam rotating LiDAR sensor modeled after the Hesai XT32 for 3D spatial mapping and obstacle avoidance.
• BEV Camera: An orthographic Bird's-Eye View (BEV) projection providing RGB imagery, dense semantic labels, and 2D bounding boxes.
• Proprioception: High-frequency IMU measurements and wheel odometry.

---

📂 Data Structure & Modalities

The raw data is generated and recorded in ROS 2 bags. Using the provided extraction pipeline (bag_extractor.py), the synchronized frames are unpacked into the following directory structure:

• rgb/ - Front-facing RGB images (.jpg).
• depth/ - Front-facing depth maps (.npy).
• lidar/ - 3D Point clouds transformed to the base link (.npy).
• bev_label/ - Orthographic semantic segmentation masks (.png).
• extrinsics/ - lidar_to_base and cam_to_base transformation matrices (.npy).

---

🏷️ Traversability-Aware Annotations

For semantic segmentation and 2D object detection across both the front and BEV cameras, we utilize classes designed specifically for agricultural navigation. Crucially, we separate the crop into traversable and non-traversable elements:

ID	Class	Description
0	ground	The mapped soil textures.
1	leaf	Corresponds to the leaf structures of the crops. This is considered a navigable (soft) class, as the robot can brush past it without causing significant damage.
2	branch	The main stem or thick branches of the crop. This is a non-navigable (hard) class that the robot must actively avoid running over.
3	weed	Corresponds to the weed assets. This class tests perception models on low inter-class separability, as weeds and crops often share visual characteristics.
4	obstacle	Encompasses the traffic cones, humans, tractor, and wild boar.

---

🔧 Utils

Besides the recorded bag files and a script to extract data from it, we also provide the base scene, the open-source assets, and the script used to generate the scenes.