Vision is Tailor's AI drone pilot. It sees the course through a single forward-facing camera, plans a path through every gate, and flies the race — entirely on its own.
8 min
Max run duration
120 Hz
Physics tick rate
0
GPS signals
1
FPV camera
System architecture
Six components, one control loop, zero human intervention.
Lightweight object detection (YOLOv8-nano class) identifies gates and obstacles from the FPV stream in under 15 ms.
Fuses IMU, attitude, and odometry into a coherent position/velocity state at 120 Hz. VIO fallback if needed.
Maintains the gate queue, tracks obstacles, and sequences navigation targets as the drone progresses through the course.
Reactive pursuit for qualification. Proportional guidance toward gate centres with obstacle avoidance.
Converts trajectory into MAVLink velocity commands at 50–120 Hz. Delegates stabilisation to the simulator.
State machine governing the run lifecycle. Handles connection, race start, gate sequencing, finish, and fault recovery.
Control pipeline
The main loop runs at 50–120 Hz. Perception is pipelined in a separate process to avoid blocking control.
MAVLink telemetry + FPV frames arrive over UDP
Gate & obstacle detection via lightweight CNN (parallel process)
Fuse IMU + odometry → state vector at 120 Hz
Compute trajectory toward next gate centre
SET_POSITION_TARGET_LOCAL_NED → simulator at 50–120 Hz
Technical spec
Key parameters from the AI Grand Prix technical specification.
Tech stack
We're assembling a team for the AI Grand Prix. If you work in computer vision, control systems, or drone autonomy — let's talk.
Join the team