Section
AI / ML
Stack
Computer Vision · YOLOv8 · PyTorch · Edge AI
Category
Computer Vision · Embodied AI
Status
Simulation complete, onboard inference optimisation in progress
01 — Challenge
The core engineering challenge
Full-precision vision models are too slow for real-time drone navigation on constrained hardware — quantisation and architecture choices need to be made as first-class design inputs.
03 — Decisions
Engineering decisions and what they cost
01INT8 quantisation reduces inference latency by 3× with acceptable accuracy trade-off on the target classes.
02Simulation-first in Gazebo before any physical flights to catch tracker failures safely.
03Tracking and detection are separate modules — detection runs at 10 fps, tracking at 30 fps.
04 — Stack
Computer VisionYOLOv8PyTorchEdge AIGazeboROS