Precision AI

The project

Precision AI build groundbreaking autonomous drones that can precisely identify the composition of a field, and spray only parts that need it, to remove weeds for example. They asked the agency to design and develop an immersive website promoting their product to both investors and future clients.

The project is live at https://www.precision.ai/

My contribution to the project

• I conceptualized the 3D experience of the landing page, first as storyboards and mood boards to share the vision to the client

• I remodeled the drone from scratch based on the reference model for a huge performance gain

• I modeled, textured, and animated the 3D scene in Blender. It was then optimized for web and exported for the developer

• We worked and iterated with the developer for weeks to achieve the best result possible in the timeframe

• We worked closely with an artistic director to design the skybox visible in the background

Challenges

The main challenges on Precision AI were about scale. Rendering one blade of grass is not an issue, but thousands can quickly become one!

Make the drone model web friendly

I started by reworking the 3D model of the drone to make it usable on the web. It came from engineers who needed an extremely detailed model to serve as a reference. Everything was modeled: from the hatches hiding electronics to the small metal parts that hold the rotors.

It was made of 2,300,000 vertices, and the mesh of the object (its structure, how the different vertices are connected to form edges and faces) was not adapted to the needs of the web. Some faces were intersecting, not attached to each other, causing small holes in the 3D model that were visible and created visual artifacts.

The quickest way to do this was to redo the model from scratch, taking the original 3D model as a reference, and adding only the necessary level of detail. This allowed us to get to less than 10,000 vertices, and to be able to display it several times in the scene without performance concerns.

Render a field full of grass and weeds

We needed to render a dense field for the drones to work on, without having too much impact on the performance.

At first, we thought of using 2D images, but the camera movements we needed would reveal the trick, as it goes from a normal view to a bird's eye view, and a top view.

We decided to create extremely simple 3D models of blades of grass, each made of a few dozen of vertices, which we instantiated and chained together randomly, to create rhythm and a more realistic, less copy pasted field. Weed models are also spread across the field.

We managed to create a field with 6 different low poly plants randomly distributed.

…And make it look endless

The end of the field mesh is hidden by a fog. It is a little levelled, just enough to hide the next instance of the field when it appears in the distance.

It is a seamless model, which is duplicated and moves on a rail towards the camera. We only have to display two of them at the same time, one after the other, to give the impression that it extends endlessly.

Bézier curves following the leveling of the field allow positioning the plants at the right height.

The skybox anchors the field in a larger universe. We can see forests and small hills behind the field.

What I learned

Precision AI gave me the opportunity to broaden my understanding of retopology and mesh optimization, as it was the first time I needed to produce meshes that would be reused and instantiate tens to thousands times.