The AI Grand Prix is hosting a competition for fully autonomous drone software, so I have decided to register and work on my own tools + autonomy model.

Architecture

Two pipelines: world generation (synthetic training data) and vision-to-control (onboard inference).

World Generator

Diffusion model outputs 3D Gaussian Splatting (3DGS) scenes conditioned on track layout graphs. 3DGS enables photorealistic rendering at arbitrary viewpoints, ~2ms rasterization, and easy domain randomization via splat perturbation.

Vision Model

Mamba SSM backbone instead of transformer: linear complexity (critical at 200 FPS), implicit memory via recurrent state, constant inference time.

Latency Budget

At 150 km/h: 4.2 cm/ms. Target 5ms loop:

Positional error stays under 21 cm—workable for ~1m gate gaps.