Edge AI Deployment Checklist for Jetson Orin [2026]

Robotics teams deploying edge AI on Jetson Orin need a repeatable checklist that covers firmware, storage, containers, model runtime, telemetry, and field security before the robot leaves the lab. As of July 01, 2026, NVIDIA's current Orin Nano ISO path is JetPack 7.2, while JetPack 6.2.1 remains important for Jetson Linux 36.4.4 fleets. Use this as a fast preflight reference for perception, autonomy, and inspection robots.

Live Search Filter

Paste this small client-side filter into an internal runbook page when your team wants a searchable deployment checklist without standing up a service.

<input id="edge-ai-filter" type="search" placeholder="Filter checks, commands, owners..." aria-label="Filter deployment checklist">
<ul id="edge-ai-checklist">
  <li data-tags="firmware jetpack qspi">Confirm Jetson UEFI/QSPI firmware before installing JetPack 7.2.</li>
  <li data-tags="storage nvme model cache">Use NVMe for models, containers, datasets, logs, and replay buffers.</li>
  <li data-tags="docker container runtime">Install Docker and NVIDIA Container Toolkit from the JetPack path.</li>
  <li data-tags="tensorrt onnx inference">Build TensorRT engines on the target software stack or matching CI image.</li>
  <li data-tags="security secure boot luks">Plan Secure Boot, encrypted rootfs, key custody, and recovery media.</li>
</ul>
<script>
const input = document.querySelector('#edge-ai-filter');
const rows = [...document.querySelectorAll('#edge-ai-checklist li')];
input.addEventListener('input', () => {
  const q = input.value.trim().toLowerCase();
  rows.forEach(row => {
    const text = `${row.textContent} ${row.dataset.tags}`.toLowerCase();
    row.hidden = q && !text.includes(q);
  });
});
</script>

Filterable Release Fields

• Board: AGX Orin, Orin NX, or Orin Nano, including carrier-board revision.
• BSP: JetPack and Jetson Linux release, including QSPI firmware path.
• Runtime: TensorRT, DeepStream, ROS 2, CUDA, and container image digest.
• Robot context: camera count, frame rate, control-loop rate, safety monitor, and remote-update channel.
• Security: Secure Boot state, encrypted storage scope, production keys, and recovery procedure.

Keyboard Shortcuts Table

Use these shortcuts in your deployment console, browser runbook, or SSH-heavy robotics workflow. The goal is to reduce mistakes during repetitive bring-up and field triage.

Commands Grouped by Purpose

Identify the Device and Release

cat /etc/nv_tegra_release
uname -a
cat /etc/os-release
dpkg-query -W nvidia-jetpack || true

• /etc/nvtegrarelease gives the Jetson Linux/L4T release marker used in fleet audits.
• dpkg-query -W nvidia-jetpack is useful on package-managed JetPack installs.
• Record outputs in CI artifacts, robot service tickets, and release notes.

Install Container Runtime

sudo apt update
sudo apt install -y nvidia-container curl
curl https://get.docker.com | sh
sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl daemon-reload
sudo systemctl restart docker

• nvidia-ctk runtime configure --runtime=docker wires Docker to the NVIDIA runtime path.
• Pin production images by digest when promoting from lab to robot.
• Keep model files and calibration assets outside mutable container layers.

Run a GPU-Enabled Container Smoke Test

sudo docker run --rm --runtime nvidia --network host \
  -v /tmp/argus_socket:/tmp/argus_socket \
  nvcr.io/nvidia/l4t-base:r36.4.0

• Use a tag that matches your Jetson Linux generation; update the example for R39.2.0 fleets when NVIDIA publishes your target base image.
• Mount camera, model, and log paths explicitly instead of relying on host-global state.
• Prefer immutable image digests for signed release candidates.

Build and Validate TensorRT Engines

/usr/src/tensorrt/bin/trtexec \
  --onnx=model.onnx \
  --saveEngine=model.plan \
  --fp16 \
  --verbose

• trtexec is the quickest sanity check before integrating inference into ROS 2 or DeepStream.
• Generate engines on matching hardware and software when precision, plugins, or dynamic shapes are involved.
• Track input shape, precision, plugin libraries, and engine build logs with the model artifact.

Monitor Power and Thermals

sudo /usr/sbin/nvpmodel -q
sudo /usr/sbin/nvpmodel -m 0
sudo jetson_clocks
tegrastats

• nvpmodel -q shows the active power mode before benchmarking.
• nvpmodel -m 0 selects mode ID 0; confirm the ID table for your exact Orin module.
• tegrastats should run during camera load, inference, planning, and actuation together.

DeepStream Container Pulls

docker pull nvcr.io/nvidia/deepstream:9.0-samples-multiarch
docker pull nvcr.io/nvidia/deepstream:9.0-triton-multiarch

• DeepStream 9.0 publishes multiarch Jetson containers for samples and Triton workflows.
• NVIDIA notes Jetson DeepStream containers are deployment-oriented; plan native or workstation builds accordingly.
• Use DeepStream for multi-camera video analytics pipelines that need GStreamer, batching, and hardware decode.

Configuration Checklist

Firmware and Install Path

• For Orin Nano Developer Kit, JetPack 7.2 uses a Jetson ISO installer path and NVIDIA says SD-card images are no longer supported for that path.
• Confirm UEFI/QSPI firmware before installing JetPack 7.2; NVIDIA documents a 36.x firmware gate for the Orin Nano flow.
• For JetPack 6.2.1, NVIDIA documents Jetson Linux 36.4.4, Linux kernel 5.15, UEFI bootloader, and an Ubuntu 22.04-based root file system.

Robot Release Manifest

robot_release:
  board: jetson-orin-nx-16gb
  carrier: custom-rev-c
  jetpack: 6.2.1
  jetson_linux: 36.4.4
  container_image: registry.example.com/perception@sha256:...
  model_artifact: warehouse-detector-2026-07-01.plan
  power_mode: production-thermal-validated
  sensors:
    cameras: 6
    lidar: 1
  rollback_slot: previous-signed-image

• Keep the manifest next to the container image and deployment bundle.
• Use TechBytes Data Masking Tool before sharing logs that include serial numbers, facility names, customer images, or map coordinates.
• Require a human-readable changelog for model, runtime, and calibration changes.

Preflight Criteria

• Boot: cold boot, warm reboot, watchdog recovery, and read-only failure path tested.
• Inference: latency, throughput, dropped-frame rate, and memory pressure measured with real sensors.
• Control: autonomy stack keeps control-loop deadlines while AI load is saturated.
• Network: OTA updates, SSH access, time sync, and offline mode are documented.
• Observability: logs, metrics, crash dumps, and model version are available after field incidents.

Advanced Usage

Secure Boot and Encrypted Storage

• NVIDIA's current Jetson Orin Secure Boot quick start references R39.2.0 and covers key generation, fuse burning, EKB preparation, UEFI key enrollment, QSPI signing, flashing, and OS installation.
• Disk encryption uses LUKS in NVIDIA's Jetson Linux documentation; test recovery before sealing devices.
• Separate developer keys, staging keys, and production keys. Treat fuse burning as an irreversible manufacturing operation.

sudo minicom -D /dev/ttyACM0 -w -c on

Model Promotion Gates

1. Export from training to ONNX with fixed metadata, dataset hash, and preprocessing version.
2. Build a TensorRT engine on the matching Jetson software stack or a validated equivalent.
3. Run replay tests with representative lighting, vibration, motion blur, and sensor faults.
4. Measure CPU, GPU, memory, temperature, power draw, and actuator-loop timing at the same time.
5. Promote only signed model bundles that can roll back independently from the base OS.

Production Rollout Pattern

git tag robot-2026.07.01-orin
cosign sign registry.example.com/perception@sha256:...
ssh robot-17 'sudo systemctl stop perception.service'
ssh robot-17 'sudo docker pull registry.example.com/perception@sha256:...'
ssh robot-17 'sudo systemctl start perception.service'
ssh robot-17 'journalctl -u perception.service -n 100 --no-pager'

• Deploy in rings: lab robot, pilot robot, limited fleet, then broad fleet.
• Keep a local rollback image for robots that lose network access after an update.
• Do not change JetPack, TensorRT engine, model weights, and camera calibration in the same release unless the test plan covers the full interaction.