Prepare your Jetson Orin Nano

• Jetson Orin Nano Developer Kit
• JetPack 7.2 or later recommended (L4T r38+, Ubuntu 24.04); JetPack 6.x also works, but some features require JetPack 7.x
• USB flash drive (16GB or larger) for the JetPack 7.2+ installer ISO
• Target storage — a PCIe NVMe SSD (recommended for sustained workloads) and/or a microSD card (64GB UHS-1 or larger)
• A computer with Balena Etcher installed to write the ISO
• A DisplayPort monitor, USB keyboard, and mouse for the install

​

Step 1: Verify board firmware

​

Step 2: Write the Jetson ISO to a USB flash drive

1. Download the Jetson ISO. Get the JetPack 7.2+ Jetson ISO (e.g. r39.2) from the JetPack download page.
2. Write it to a USB flash drive. Open Balena Etcher, select the ISO, select your USB flash drive (16GB+), and click Flash.

​

Step 3: Install your target storage

• NVMe SSD (recommended): install it into the M.2 Key M slot on the underside of the carrier board and secure it with the screw.
• microSD card: insert it into the slot on the module underside.

​

Step 4: Boot the installer and install Jetson Linux

1. Connect a DisplayPort monitor, USB keyboard/mouse, and the installer USB flash drive, then power on.
2. Press Esc to enter UEFI, open the Boot Manager, and select the USB drive.
3. When prompted to update the QSPI firmware, press Y within 30 seconds to confirm. Allow any UEFI capsule updates to finish — the board may reboot, which is normal.
4. Select your target storage (NVMe SSD or microSD card) and confirm. The installer writes Jetson Linux and handles partitioning and boot configuration automatically — no manual dd, extlinux, or partition-resize steps are needed.
5. When it finishes, reboot from the target storage and complete initial setup (accept the EULA, choose language/keyboard/timezone, connect to the network, create your user account).

df -h /

​

Step 5: Configure headless operation

sudo systemctl set-default multi-user.target
sudo systemctl enable ssh

​

Step 6: Configure WiFi for headless boot

sudo nmcli connection modify <your-wifi-name> connection.autoconnect yes
sudo nmcli connection modify <your-wifi-name> connection.autoconnect-priority 10
sudo nmcli connection modify <your-wifi-name> connection.permissions ""

​

Step 7: Configure 5G modem failover (optional)

nmcli device status

sudo nmcli connection modify "Wired connection 2" connection.autoconnect yes
sudo nmcli connection modify "Wired connection 2" connection.autoconnect-priority 5
sudo nmcli connection modify "Wired connection 2" connection.permissions ""

ip route

​

Reboot and verify

sudo reboot

​

Step 8: Configure Ethernet for robot connections (if applicable)

ip a

sudo nmcli connection add type ethernet ifname <interface-name> \
  con-name robot-link ip4 192.168.123.11/24
sudo nmcli connection up robot-link

ping 192.168.123.161

​

Step 9: ML worker image (JetPack 7.2+)

​

Supported hardware

​

Requirements

• JetPack 7.2 or later (L4T r38+, Ubuntu 24.04) on any Jetson Orin SKU. The GPU-accelerated -jetson ML worker image requires JetPack 7.x — on JetPack 6.x hosts the edge stack still runs, but this worker falls back to CPU. The -jetson image ships CUDA 12.9 runtime libraries; the JetPack 7 CUDA 13.x driver (injected at container start) is backward-compatible with them.
• nvidia-container-toolkit installed and enabled (shipped by default with JetPack via apt install nvidia-jetpack).
• Roughly ~9 GB of free space for the first pull. The image bundles PyTorch, ONNX Runtime GPU, CUDA 12.9 runtime libraries, and cuDNN so the container works on any JetPack 7.2+ Orin host without needing CUDA installed system-wide.

​

What gets selected

• Pulls the -jetson image variant automatically (instead of the base or GPU variant).
• Passes --runtime nvidia -e NVIDIA_VISIBLE_DEVICES=all -e NVIDIA_DRIVER_CAPABILITIES=all so the nvidia-container-runtime injects the Tegra iGPU device files and the CUDA driver into the container.
• Skips Hailo and discrete-GPU device passthrough.

​

Verifying it’s using the Jetson image

docker ps --format '{{.Image}}\t{{.Names}}' | grep edge-ml-worker

docker exec "$(docker ps -q -f name=cyberwave-worker)" \
  python -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"

​

Overriding the auto-selection

{
  "CYBERWAVE_WORKER_IMAGE": "cyberwaveos/edge-ml-worker:local"
}

​

Troubleshooting

• Worker starts on CPU instead of -jetson — check that /etc/nv_tegra_release exists on the host and that docker info lists the nvidia runtime. If both are present but the edge stack still picks CPU, look for a falling back to warning in the edge service logs — the -jetson image probably failed to pull.
• torch.cuda.is_available() returns False inside the worker — the container is running on CPU. Confirm NVIDIA_VISIBLE_DEVICES=all is set in the container env (docker inspect <container> | grep NVIDIA), that the container was started with --runtime nvidia, and that nvidia-smi works from the host.
• Out of memory when loading a YOLOE model — the model is too large for this Jetson SKU. Consult the sizing table above; either downshift to a smaller variant (yoloe-11n / yoloe-11s) or move the workload to Orin NX / AGX.
• Very slow first start — the image is ~3–4 GB and Jetson NVMe throughput on first pull can be modest. Subsequent restarts reuse the local layer cache.