When the Jetson Nano launched I wanted to put one in a moving vehicle and see what I could do with a single camera. I designed a small RC chassis, printed it, wired up the motor controller and camera, and built a perception stack around it.
Perception on a budget
The honest reason there’s no LiDAR in this project is that I couldn’t afford one. Everything the car “sees” comes from a single camera. Obstacle detection is YOLO-based segmentation — not great, but adequate for the scale and speed of a toy car in a room. The car can identify the ground plane, flag obstacles within a forward cone, and plan avoidance manoeuvres that favour the widest gap.
What I learned
This was the first time I was building a system where the interesting bugs lived in the ML layer rather than the code layer. When the car started driving into a dark rug it wasn’t a control bug; it was the segmenter mis-labelling the rug as “not ground.” Debugging that meant thinking about training data and lighting invariance — which I’d read about, but it’s different when it’s moving around your living room.
Not a product. A proof that I could pull computer vision off a GPU and into something that moves around, and that I could rebuild an RC chassis from CAD without breaking the geometry.