Skip to main content

Edge Deployment

Waldo can push trained models to remote inference endpoints — small Linux boxes running close to the cameras. Two reference targets are documented here.

NVIDIA Jetson (Orin Nano / NX / AGX)​

Jetson devices run Ubuntu (L4T) and have a CUDA-capable GPU. The Waldo edge image is a slimmed compose stack: just app (in lightweight mode) + a single inference worker.

# On the Jetson
git clone https://github.com/oldhero5/waldo.git --depth 1
cd waldo/edge/jetson
docker compose up -d

Register the device with the central API so it shows up on the Deploy page:

curl -X POST https://waldo.example.com/api/v1/devices \
  -H "Authorization: Bearer $WALDO_API_KEY" \
  -d '{"name": "front-gate-jetson", "kind": "jetson-orin-nano", "ip": "192.168.1.42"}'

Heartbeats are sent every 30 seconds.

Raspberry Pi 5 + Coral USB TPU​

For lower-power deployments, export the YOLO model to TFLite Edge TPU format:

curl -X POST https://waldo.example.com/api/v1/models/$MODEL_ID/export \
  -H "Authorization: Bearer $TOKEN" \
  -d '{"format": "tflite_edgetpu"}'

On the Pi:

# Coral runtime
sudo apt install -y libedgetpu1-std
# Pull the edge runtime image (matches the central Waldo version)
docker run -d --device /dev/bus/usb \
  -e WALDO_API=https://waldo.example.com \
  -e WALDO_API_KEY=... \
  -e DEVICE_NAME=garage-pi \
  ghcr.io/oldhero5/waldo-edge:pi5-coral

OTA model updates​

When you promote a new model in the central UI (or call POST /models/{id}/promote), every connected edge device picks up the new version on its next heartbeat and downloads the weights from MinIO via a presigned URL.

Logs from each device are pushed back via POST /devices/{id}/sync-logs so you can audit edge predictions centrally.