My background is in autonomous vehicles and aerospace, with a foundation in computer vision. When I started looking at agriculture, I saw an environment where autonomy could have real, immediate impact. Farms needed more output with less labor, and autonomy was the clearest path to get there. That’s what led me to co-found Bonsai.
Agriculture Is a Challenging Environment. That’s What Makes It Valuable.
Why outdoor autonomy complexity is a feature, not a problem
A lot of people assume agriculture is a straightforward domain for robotics. Open fields, slow speeds, predictable conditions. In practice, it’s quite different.
The environment changes constantly. Plants grow and alter their appearance season over season. You have weather, dust, heat and debris. Terrain varies across farms, geographies and crop types. And there’s always seasonality, which creates real time pressure on operations.
But that complexity is actually what makes agriculture such a good starting point. If you can build an autonomous system that works reliably here, you have built something that can generalize. You’ve solved the hard problem in a real environment. And from there, you can expand.
That is the core of how we think about it at Bonsai. Start in a difficult, unstructured, real-world environment. Build something that actually works in it. Then use that foundation to scale into other industries.
Why Agriculture Made More Sense Than Autonomous Vehicles or Aerospace
The practical advantages that make ag the right starting point for physical AI
Before Bonsai, I spent time in the autonomous vehicle space, working at Mercedes-Benz, Volkswagen and Toyota. The domain is genuinely complex. Pedestrians, cyclists, unpredictable drivers, dense intersections. The edge case space is very large, and the regulatory environment to match.
Aerospace looked simpler at first. Closed airspace, fewer unpredictable variables. But flying over populated areas brings its own regulatory burden, and certification requirements add up fast.
Agriculture offered a different set of conditions:
- Reduced Regulatory Complexity: Operations occur primarily on private land.
- Simplified Safety: Reduced human presence minimizes traditional safety risks.
- Immediate Productivity Impact: Addresses urgent labor shortages and rising input costs.
- Global Food Security: Direct impact on food production, which matters to everyone.
The environment is complex enough to build genuinely robust systems. But the constraints that make autonomous vehicles and aerospace so difficult are largely absent. That combination made agriculture the right place to start.
Data Is the Foundation of Everything
Why real-world deployment is the only way to build a system that actually works
At Bonsai, our approach from the beginning has been to deploy early and learn from each deployment. We didn’t start with a finished product. We started with sensors, installed them on existing machines, let operators drive and collect data passively. That was step one.
The reason we prioritized this is simple. Data is the most valuable asset in autonomy development. Without data, you can’t build models. Without models, you can’t have autonomy.
Making autonomy work requires three components operating together in a continuous loop:
- Perception: understanding where you are relative to your environment and identifying what’s around you
- Decision making: taking in what you perceive and forming a plan
- Controls: translating that plan into actual machine commands, steering, speed and actuation
You can simulate some of this, and simulation is genuinely useful for iterating quickly. But simulation alone won’t give you the corner cases and real-world variability that only come from actual field operations. The grounding to reality has to come from real deployments.
Every deployment generates learnings. Those learnings feed back into the model. A better model enables better deployments. More deployments generate more data. That is the flywheel we have been building since day one, and it’s what allows the system to keep improving over time.
A System Designed to Generalize, Not Specialize
The benefits of a fully integrated autonomy stack
From the start, we made a decision to build a fully integrated system rather than combining third-party components. The reason is practical. Most third-party solutions are built to solve one specific piece of the problem, and they’re not designed to work together in a generalized way.
Autonomy is a systems problem. Perception, decision making and controls need to function as a unified whole. When you piece together solutions from different providers, the overall system becomes harder to improve and harder to scale, especially in the variable conditions that outdoor environments produce.
Our system is built to run across different vehicle types, terrains, crops and geographies from a single generalized model. When we integrate with a new vehicle, we need to understand the sensor configuration, the dimensions and the kinematics. From there, the system adapts. We don’t rebuild for each new application.
We also started with monocular cameras rather than a complex sensor array. The reason is simplicity. A single camera, with the right software, can reconstruct the surrounding environment in 3D. Simpler hardware is more robust, easier to maintain and more practical to deploy at scale. We add sensors when a specific application calls for it, but the default is to keep the system as lean as possible.
What This Enables Beyond Agriculture
How solving outdoor autonomy in ag opens the door to construction, mining and more
The reason we focused on agriculture isn’t only the impact on food production. It’s that the core challenges of outdoor autonomy are shared across many industries.
Whether you are operating in an orchard, on a construction site or in a mine, the fundamental problems look similar:
- Navigate reliably without depending solely on GPS
- Identify and respond to obstacles in real time
- Handle variable lighting, dust and unstructured terrain
- Generalize across different environments without retraining from scratch
A system that can do these things in agriculture can do them in other off-road domains as well. The model doesn’t distinguish between an almond orchard and a mining site at a fundamental level. It perceives its environment, makes decisions and executes. The domain changes. The capability transfers.
That’s why we started here. Agriculture is a productive, demanding environment that pushes systems to perform reliably under real conditions. Building something that works here has provided us with the strong foundation we are now using to expand into other industries where the same challenges exist.
Ugur Oezdemir is the CTO and Co-Founder of Bonsai Robotics