Lycan Automotive

The problem

Lycan was building consumer EVs for Indian roads. Unstructured traffic, unmarked lanes, and pedestrians who negotiate rather than wait. The benchmarks trained on Californian highways do not transfer.

I built the perception and planning stack that had to survive it, across CARLA, Highway Env, and a custom simulator, trained on a mix of our own data and KITTI and UCB.

3D object detection

Single-stage and multi-stage monocular 3D detection in PyTorch and Numba, covering vehicles, pedestrians, cyclists and trucks, at 87% mAP from a single camera and no lidar.

Lane detection

84% accuracy on real-time camera feed, on roads where the lane markings are frequently a suggestion.

Collision avoidance

A planner in Highway Env that avoids vehicles and pedestrians at a 95% success rate.

Outcome

I assembled the pieces into a self-driving proof of concept in a custom simulator. It secured the company’s first funding round.