Deploying Autonomy on the Farm with Tyler Niday of Bonsai Robotics

With something like this, I imagine the technology continues to get better and better and better and better.

How do you make sure that that's a competitive advantage for you rather than just like inviting all sorts of competition?

Yeah, most definitely.

So it's an interesting time, right? Where, you know, at Blue River, I remember when in 2015, the first NVIDIA deep learning chip came out, right? And we had YOLO bounding boxes, and it was the most amazing thing ever doing green on green sea and spray.

And now we're at the point where you could literally deploy an AI model with no code from what you perceive in your environment to a perfect 3D world.

I think we're at a really interesting point where data is kind of the kingmaker in this space right now.

And one thing we've always been super keen on is huge amounts of data, of the whole machine data, of every interaction the human makes with the machine.

50 million samples of mcaps which are a robotics term for um like all the data in one thing that we're training these big models on so what we're able to do with that is instead of writing lines and lines and lines of deterministic code for one specific application like blast spring and orchard we have no code from the camera images input to a scaled 3d world output so we can go into a strawberry field and open you know corn cotton or soy field an almond field, a macadamia nut field and have no code and deploy this AI model that works and perceives as good as a human without LIDAR, without stereo vision.

So that's, I think, the big thing that we're really keen in and excited about is being able to go across the TAM to all these diverse applications with no changes to the AI system.

Is that where the next frontier is in terms of training the model? A lot of times I'll hear like that, like we have all this data, we can train the models to be better.

And I often find myself wondering, like, how much better does a model need to be? Or is it just to kind of reach new use cases?

Yeah, so I think historically so much of robotics and agriculture have been like a deterministic case where we have a 2D, you know, segmentation map or model, right? Where we detect things in an image like, hey, this is a tree, this is an obstacle, this is a human, and that's all in 2D, right? And then we have some way to convert it to 3D space to actually navigate, whether it's a LIDAR or stereo vision depth maps.

The struggle with those sensors is if there's dust, if there's debris, we're perceiving in 3D.

What we can do now is just start in 3D.

A human doesn't look at an obstacle and be like, oh, this is a 2D image, and then I need to create the 3D world.

We have an AI model on our head that says like, has all these associated priors to this model.

So we know roughly where it is and what it can be, right? So that's what we're able to do now.

And I think that's the big advantage.

A lot of ag is all written on historic deterministic stacks that can do one thing.

And then you write hundreds of thousands of lines of code to do more.

And now we don't have to do it, which is

pretty crazy.

That's really cool.

And so is the vision for the company to kind of continue to partner with these OEMs and just kind of bring this technology to more agricultural use cases or kind of help us understand where all this leads?

Completely.