Where to look¶

The repository¶

Path	What it holds
`examples/`	Complete apps, each laid out as `controller/` (STM32), `gap8/` (GAP8) and `host/` (your computer), with a README
`lib/`	The C libraries for the GAP8 and STM32; `lib/README.md` lists every component
`src/cf_lab/`	The Python package and the `cf-lab` command
`templates/`	Starting points: the smallest out-of-tree controller and the smallest neural network
`configs/drone.yaml`	The drone's radio address, mocap, camera WiFi and telemetry settings
`external/`	Pinned submodules: crazyflie-firmware, the AI-deck ESP32 firmware, NEMO and DORY. `external/README.md` says which are forks; each fork's `CHANGES.md` lists what it changes
`docker/`	The images for GAP8 builds and for NEMO/DORY
`tests/`	The Python tests: `pytest -n auto`
`docs/`	This site

Examples¶

Example	What it shows	Parts
`idle`	Minimal STM32 app and a parked GAP8: the starting template	controller, gap8
`telemetry_demo`	GAP8 and STM32 counters over CPX, recorded on the host	controller, gap8, host
`debug_demo`	Source-tagged console logging from the STM32 and GAP8	controller, gap8, host
`telemetry_bench`	Benchtop stress test of the telemetry recorder	controller, gap8, host
`camera_stream`	The AI-deck camera as a WiFi video server, viewed and recorded with mocap	gap8, host
`camera_bench`	Regression bench for camera_stream: builds, flashes and measures every camera mode	gap8, host
`wifi_bench`	Regression bench for the AI-deck WiFi link, GAP8 to ESP32 to host, over TCP and UDP	gap8, host
`mocap_positioning`	Propellers-off live mocap position indicator	controller, host
`mocap_waypoint`	Flying mocap waypoint demo on the stock PID controller	controller, host
`hover_neural_network`	PyTorch hover policy run at 100 Hz on the GAP8 alongside JPEG video	controller, gap8, host, model
`nn_bench`	Whole DORY networks in GVSOC and on the GAP8, every layer checked against the host	gap8, host
`scheduler_bench`	DORY networks sharing the GAP8 with the camera and the WiFi video stream	gap8, host
`kernel_bench`	GAP8 quantized-kernel development and profiling against PULP-NN	gap8, host
`jpeg_cluster_bench`	Verification bench for the GAP8 JPEG encoder, FC and cluster	gap8, host
`hyperbus_bench`	AI-deck HyperFlash, flash filesystem and HyperRAM checks	gap8, host
`flash_bench`	Regression bench for GAP8 flashing over the radio	gap8, host

Each example's README has its build, flash and run steps. Run the examples from the repository root.

The Python package¶

Module	What it does
`cf_lab.config`	Loads `configs/drone.yaml`
`cf_lab.backends`	The radio link to the drone, and an in-memory fake for dry runs
`cf_lab.crazyflie`	Firmware helpers: parameters, log variables, the boot self-test, the IMU
`cf_lab.runner`	Building blocks for an example's `host/run.py`: kill switch, watchdog heartbeat, sessions
`cf_lab.mocap`	Motion-capture pose sources (OptiTrack over NatNet, or a fake) and the stream to the drone
`cf_lab.camera`	The AI-deck image stream on the host: the feed, the viewer, rendering video, finding the deck
`cf_lab.telemetry`	Records a run's streams to Parquet, then loads and checks them
`cf_lab.analysis`	Turns recorded runs into datasets; checks clock alignment
`cf_lab.preflight`	Read-only checks before flying
`cf_lab.sim`	Differentiable, batched vehicle simulation
`cf_lab.aideck`	Compiling a model for the GAP8 and flashing the AI-deck

Neural networks¶

cf-lab compile-model turns a PyTorch model into C for a GAP8 app: NEMO quantizes it and DORY generates the code, both in Docker. Start from templates/neural_networks/README.md; examples/hover_neural_network is a complete example.