Source-linked AI summary
AgriGen: Large-Scale Scene Generation Framework for Photorealistic Agricultural Robotics Simulation
Utkarsh Bajpai, Serge Tleiji, Cédric Pradalier, Stéphanie Aravecchia
TL;DR
Agricultural robotics evaluation depends on access to diverse field conditions, while agricultural simulation has lacked the scale and environmental diversity available in other robotics domains. AgriGen uses procedural generation and runtime asset streaming to create large-scale agricultural simulations; it supports multiple crop categories and runs with performance tied to local scene complexity.
Problem
Agricultural robotics needs diverse, controllable field conditions for evaluation, but existing simulators have generally relied on small, simplified scenes with limited environmental diversity.
Source: §I · p. 1Method
AgriGen combines procedural terrain and vegetation generation with runtime asset streaming in a ROS-enabled Isaac Sim framework supporting row crops, orchards, and vineyards.
Source: §II · p. 1, C. Asset Streaming · p. 2, §III · p. 2Results
AgriGen ran successfully on all three tested hardware platforms, with FPS scaling with GPU capability and performance depending on local scene complexity rather than total environment size.
Source: B. Performance Evaluation · p. 3Takeaways & Limitations
AgriGen supports systematic and reproducible evaluation across diverse, large-scale agricultural environments through controlled field-condition variation and access to ground-truth annotations.
Source: §IV · p. 4
Abstract
from arXiv · showhide
Agricultural robotics is advancing rapidly, yet progress remains constrained by limited field access, lack of control over field conditions, geographic variability, and seasonal crop cycles. These factors make it difficult and costly to acquire diverse agricultural datasets, resulting in limited evaluation and reduced system robustness. While other robotics domains have scaled learning and evaluation through high-fidelity simulation, agricultural robotics still lacks comparably capable tools. In this paper, we present a ROS-integrated framework, built on Isaac Sim, for large-scale procedural generation of agricultural environments. The framework supports photorealistic rendering, physics simulation, and domain randomization at scales relevant to robotics research, with built-in support for row crops, orchards, and vineyards and straightforward extensibility to additional crop categories. Project Page: https://baj31415.github.io/agrigen/
I. INTRODUCTION
Agricultural robotics evaluation needs diverse field conditions and control over them, but comparable large-scale agricultural simulation remains limited. AgriGen addresses this gap with a framework for large-scale agricultural robotics simulation.
I. INTRODUCTION
- Robust evaluation requires varied crops, terrain, and illumination, while data-driven agricultural robotics also needs large labeled datasets.Field access and control over layout, crop density, and weed distribution are typically required.Source: §I · p. 1
- Earlier agricultural simulators used simplified environments, limited photorealism, coarse assets, or small scene scales, [8],.Later Unity- and Unreal-based approaches improved visual realism but remained limited in scale and environmental diversity,.Source: §I · p. 1
- AgriGen targets realistic field-scale simulation with diverse vegetation and terrain, enabling settings that better capture real agricultural environments.The paper identifies existing frameworks as lacking this combination of scale, vegetation diversity, and terrain variation.Source: §I · p. 1
II. OUR APPROACH
AgriGen procedurally generates large-scale agricultural environments through terrain and vegetation instantiation followed by runtime asset streaming.
II. OUR APPROACH
- AgriGen structures simulated environments through terrain and vegetation instantiation, then progressively loads terrain and vegetation assets at runtime.The paper presents these as three phases of its generation process.Source: §II · p. 1
- Terrain and vegetation instantiation form the environment before streaming progressively loads their assets into the simulator.The process is designed for large-scale agricultural environments.Source: §II · p. 1
- The framework’s generation process is procedural and flexible, with Fig. 1 providing an overview of the method.The overview covers terrain instantiation, vegetation instantiation, and runtime streaming.Source: §II · p. 1
A. Terrain Instantiation
AgriGen models field terrain with a procedural elevation field and surface appearance to represent variation from broad landforms to fine soil roughness.
A. Terrain Instantiation
- Perlin noise generates an elevation field combining broad hills and valleys with smaller mounds and ridges.The elevation map is converted into a triangular mesh representing the ground surface.Source: A. Terrain Instantiation · p. 1
- The procedural terrain method combines elevation geometry with mapped surface appearance to produce varied agricultural field terrain.The elevation field uses macro and meso spatial scales, while texture and bump mapping represent fine-scale surface variation.Source: A. Terrain Instantiation · pp. 1–2
- Texture and bump mapping assign terrain color, material appearance, and fine-scale surface roughness such as soil irregularities.These surface treatments add visually plausible variation representative of agricultural fields.Source: A. Terrain Instantiation · p. 2
B. Vegetation Instantiation
AgriGen procedurally distributes vegetation assets to represent crops and other elements of agricultural landscapes.
B. Vegetation Instantiation
- The vegetation framework represents crops alongside grasses, weeds, shrubs, and rocks.These assets capture vegetation and other elements found in realistic agricultural environments.Source: B. Vegetation Instantiation · p. 2
- Procedural distribution uses different sampling strategies and stores the resulting assets on a vegetation map.This map is defined over the same planar domain as the elevation map.Source: B. Vegetation Instantiation · p. 2
- The vegetation map shares its planar domain with the elevation map, linking vegetation distribution to the terrain representation.The framework uses procedural instancing to distribute vegetation assets.Source: B. Vegetation Instantiation · p. 2
C. Asset Streaming
AgriGen streams nearby terrain and vegetation tiles to support large-scale simulation, while its broader framework covers multiple crop categories and domain randomization.
C. Asset Streaming
- Streaming loads nearby tiles and unloads distant ones, keeping simulation feasible despite the computational cost of thousands of vegetation assets.The pipeline supports RTX rendering, PhysX dynamics, collision detection, and ROS 2 sensor publishing at each simulation timestep.Source: C. Asset Streaming · p. 2
III. AGRIGEN AS AN AGRICULTURAL ROBOTICS SIMULATOR
- The framework supports row crops, orchards, and vineyards, and its pipeline randomizes terrain, assets, asset distribution, and lighting.The released vegetation asset dataset contains 65 3D models, including crops, weeds, grass, and shrubs.Source: §III · p. 2
A. Simulated Sensors and ROS Integration
AgriGen publishes synchronized sensor outputs and annotations from simulated agricultural scenes, including imagery, labels, and point clouds.
A. Simulated Sensors and ROS Integration
- The robot’s egocentric view provides synchronized RGB, semantic annotation, and LIDAR point-cloud outputs.The simulator can also publish depth, IMU, and GPS data.Source: A. Simulated Sensors and ROS Integration · pp. 2–3
- Cabbage-field images can include bounding boxes, semantic labels, and instance labels for crops and weeds.Source: Figure 3 · p. 3
B. Performance Evaluation
The evaluation reports successful simulation across three hardware platforms and finds that performance depends on local scene complexity rather than total environment size.
B. Performance Evaluation
- All three tested hardware platforms ran the simulation successfully, with FPS scaling with GPU capability and depending on local scene complexity rather than total environment size.The streaming architecture loads a fixed number of tiles as the robot moves, with only minor FPS fluctuations as its viewpoint changes.Source: B. Performance Evaluation · p. 3, Table I · p. 3
C. Assets
AgriGen provides detailed plant assets for row crops, orchards, and vineyards, with an open-source dataset that can be extended.
C. Assets
- The asset collection covers row crops, orchards, and vineyards, with ten model variations per crop type, from apple and lemon trees to cabbages and sugar beets.Fig. 4 shows representative models and close-ups of their geometric detail.Source: C. Assets · p. 3
- The asset dataset is released open-source with AgriGen.Source: C. Assets · p. 3
- The framework is not limited to the listed models and can be extended.Source: C. Assets · p. 3
D. Domain Randomization
AgriGen demonstrates diverse in-field conditions through domain randomization across terrain, assets, lighting, and physics, using multiple crop types.
D. Domain Randomization
- AgriGen demonstrates the ability to produce diverse in-field conditions commonly encountered by agricultural robots through domain randomization.Source: D. Domain Randomization · p. 3
- The asset examples cover apple trees, lemon trees, cabbages, sugar beets, and grapevines, with close-up views of three crop types.Source: Figure 4 · p. 3
- Table II is identified as reporting AgriGen asset statistics.Source: Table II · p. 3
- Randomization spans terrain geometry and type, asset geometry, type and configuration, illumination, and physics parameters such as friction.Examples vary terrain morphology, planting patterns, vegetation density, ground textures, and lighting.Source: D. Domain Randomization · p. 3
IV. CONCLUSION
AgriGen combines procedural environment generation and streaming with synchronized sensor outputs for large-scale agricultural simulation. The paper reports competitive frame rates on consumer-grade hardware while preserving photorealism, and demonstrates qualitative domain randomization across field conditions.
IV. CONCLUSION
- AgriGen combines procedural terrain instantiation, vegetation distribution, and streaming-based asset management to generate photorealistic environments at scale.It provides synchronized RGB and depth images, image annotations, point clouds, and IMU and GPS measurements as ROS 2 topics.Source: §IV · p. 3
- The framework achieves competitive frame rates on consumer-grade hardware while preserving photorealism in large-scale environments.Source: §IV · pp. 3–4
- AgriGen qualitatively demonstrates domain randomization over terrain, vegetation, textures, and lighting to simulate diverse in-field conditions.The conclusion states that researchers can systematically and reproducibly evaluate systems across large-scale environments with controlled field-condition variation and ground-truth annotations.Source: §IV · p. 4, Figure 5 · p. 4