Agriculture Robotics and Edge AI
Aigen Advances Sustainable Farming With Solar-Powered Robots and NVIDIA Cosmos
Aigen
Objective
Aigen develops autonomous, solar-powered robots that mechanically remove weeds, helping farmers protect crops while reducing reliance on herbicide applications and diesel-powered field passes. Because agricultural fields change rapidly, collecting and labeling enough real-world data to train reliable AI is slow and expensive.
To overcome this bottleneck, Aigen built Alchemy, an outdoor world model that generates realistic synthetic variations in crops, weeds, soil, lighting, and depth. Alchemy includes a post-trained NVIDIA Cosmos customized for outdoor, off-road autonomy. Aigen’s synthetic data enables autonomous field weeding while saving critical time and resources. Aigen’s newest robots use three NVIDIA Jetson Orin NX modules for energy-efficient perception, navigation, and mechanical control.
Key Takeaways
99% Synthetic Data With Cosmos
- Aigen built Alchemy, its custom outdoor world model trained with 99% synthetic data, by post-training NVIDIA Cosmos 3 for more realistic and diverse agricultural environments.
Reducing Compute Utilisation to 30% With NVIDIA Jetson
- Moving onboard compute to NVIDIA Jetson Orin NX cut compute utilization from roughly 100% to about 30%, improving stability and freeing capacity for additional AI workloads without adding hardware or exceeding the robot's solar-power budget.
Efficient Hardware, Smarter Models
- NVIDIA Jetson — Expanded robot capability while reducing onboard hardware from five compute devices to three NVIDIA Jetson Orin NX modules, enabling a four-arm design and more complex robotic actions.
- NVIDIA Cosmos — Cut the time and cost of expanding training coverage by replacing months of manual field-data collection and labeling with synthetic scenario generation.
Scaling Chemical-Free Weed Control
Advancing Chemical-Free Weed Control Across Unpredictable Fields
In the realm of physical AI, agricultural robotics faces some of the most complex and uncontrolled visual environments. Across a single growing season, a soybean field undergoes a total transformation—from initial planting to full canopy—while factors like lighting, soil conditions, and varying weed pressure continuously alter the scene for autonomous systems.
The traditional approach to agricultural AI relies on years of manual data collection and annotation, yet this method struggles to scale. Because every field is unique and visual data is often limited to brief seasonal windows, capturing and labeling the full breadth of field variation becomes prohibitively expensive and time-consuming.
The impact of weeds is immediate: invasive weeds diminish harvest productivity, and managing these weeds across vast acreage presents a significant financial burden. The emergence of herbicide-resistant varieties is a growing challenge for many farmers, yet relying on human labor for removal is neither scalable nor cost-effective. Furthermore, chemical drift poses risks to personnel and adjacent vegetation, while runoff threatens local aquatic ecosystems. By deploying fully autonomous, solar-powered robotics for mechanical weed control, Aigen empowers growers to safeguard their yields while minimizing dependency on traditional herbicides and diesel-intensive field operations.
The central challenge was the gap between rapidly changing agricultural environments and the slow pace of manual data collection and labeling. To capture different crop types, growth stages, weed density, lighting conditions, and disease scenarios, the Aigen team leveraged synthetic data generation to expand training coverage. This enabled the team to train perception models to generalize to crop types, weed species, and field conditions not represented in the post-training dataset.
Developing Agricultural AI With NVIDIA Cosmos
Scaling Agricultural Training Data With NVIDIA Cosmos
With 275 herbicide-resistant weed species documented globally, seasonal field data collection alone cannot provide the visual diversity agricultural AI requires. Aigen uses Alchemy and NVIDIA Cosmos 3 to generate and label synthetic agricultural scenarios. In earlier work, Aigen post-trained Cosmos Transfer 2.5, using about 3,000 fleet-captured RGB and depth clips—roughly 8.3 hours of soybean, cotton, and tomato footage spanning different weed densities, growth stages, and lighting conditions.
The post-trained model improved results in unseen environments, including:
- More species-appropriate leaf shapes and colors, greater weed diversity, more detailed soil textures, more consistent outdoor lighting and shadows, and stronger alignment with depth maps.
- Scene descriptions using crop type, camera angle, and field conditions, producing labeled examples for perception-model training.
Building on this work, Aigen post-trained a policy on action-conditioned driving data, enabling the company to generate one million hours of action-video data in months, providing a scalable source of paired training data for future policies
Using synthetic training data from Alchemy, Aigen achieved autonomous field weeding with a perception model trained on 1% real-world data and 99% synthetic data. The approach reduced the time and costs associated with field-data collection and manual annotation. As the work expands from perception to action, the next challenge is running increasingly capable models within each robot’s solar-power budget.
“Alchemy, along with Cosmos, allows us to stand up production weed control and field analytics models so quickly that we can now deploy in the same season we see a crop. Normally that takes three to four years of collecting data, labeling it, and training models. The labels come with the data by construction rather than from a manual annotator, so we're no longer beholden to seasonality, model turnaround is hours instead of seasons, and efficient models now deploy straight to NVIDIA Jetson Orin NX on our solar-powered Element robots.”
Richard Wurden
Co-founder and CTO , Aigen
“Cosmos 3’s pretrained priors and extensible architecture provide an excellent foundation for building a custom world model. This helped accelerate Alchemy’s development, which unlocks new capabilities in outdoor unstructured autonomy.”
Usman Khan
Senior Data Scientist, Aigen
Expanding Robot Capability With NVIDIA Jetson
Building a Four-Arm Robot With Three NVIDIA Jetson Modules
Replacing five system-on-chip (SoC) devices with three NVIDIA Jetson Orin NX modules gave Aigen more onboard AI capacity with reduced power needs The additional headroom supports a four-arm design, increasing weeding throughput while keeping real-time perception, vision-based navigation without GPS, and mechanical control on the robot.
During development testing, Aigen reports that the Jetson-based architecture operated at approximately 30% compute utilization, compared with 100% on its previous system. The remaining capacity improves system stability and gives Aigen room to add capabilities beyond weed control without adding compute hardware or exceeding the robot’s solar-power budget.
Scaling Chemical-Free Farming From Field Trials to Fleet
Aigen plans to use the available compute headroom to add capabilities beyond weeding without redesigning the hardware or exceeding its solar-power budget. That flexibility is central to Aigen’s development model, allowing new software capabilities to build on the same edge-compute foundation as the fleet expands. As it broadens support across crops and field conditions, Aigen is adopting NVIDIA Cosmos 3 as Alchemy’s primary base model to expand its synthetic-data capabilities. This combination of agricultural AI and solar-powered robotics gives Aigen a scalable path toward weed control that reduces reliance on herbicides and diesel-powered field passes. Together, Aigen’s agricultural AI and solar-powered robots are helping scale chemical-free weed control and build a more sustainable future for farming.
“Aigen is the only company doing fully-autonomous, solar-powered mechanical weed control, and every watt of compute counts. Switching to NVIDIA took us down from five SoCs to three and gave each robot more capability. That’s what lets our four-arm robot run full edge AI on solar, navigate without GPS, and keep compute margin for capabilities beyond weed control.”
Richard Wurden
Co-founder and CTO , Aigen
Explore how NVIDIA Cosmos helps developers develop Physical AI faster with leading World Foundation Models, open data processing, training and evaluation frameworks.