Jetson Orin Edge AI Deployment Checklist [2026 Guide]

Deploying edge AI on Jetson Orin is less about one perfect image and more about repeatable checks: board support, power modes, containers, model runtime, telemetry, rollback, and field diagnostics. As of June 11, 2026, NVIDIA JetPack 7.2 brings Jetson Linux 39.2 support to the Orin family, while JetPack 6.2.1 remains the common production baseline for many Ubuntu 22.04 robotics stacks.

• JetPack 7.2 pairs Orin support with Jetson Linux 39.2, CUDA 13.2.1, and TensorRT 10.16.2.
• JetPack 6.2.1 uses L4T 36.4.4, CUDA 12.6, TensorRT 10.3, and Ubuntu 22.04.
• Pin power mode, thermal policy, container digest, model artifact, and camera firmware before field tests.
• Use tegrastats, Nsight, and application latency probes together; no single metric explains robot behavior.

Version Baseline

Release checklist

• JetPack 7.2: choose when you need the newer Jetson Linux 39.2, Ubuntu 24.04 base, CUDA 13.2.1, or Orin support in the JetPack 7 line.
• JetPack 6.2.1: choose when your fleet depends on Ubuntu 22.04 packages, L4T 36.4.4, validated camera drivers, or existing ROS 2 images.
• TensorRT: export engine metadata with precision, input shape, calibration set hash, GPU/DLA target, and builder version.
• Docker: pin image digests, not floating tags; NVIDIA notes Docker compatibility fixes in L4T 36.4.4.
• Security: plan signing, disk encryption, secrets rotation, and the HSM boot-image signing path before pilot builds.

Live search JS filter

Drop this filter above long command tables so field engineers can type power, camera, logs, or rollback and narrow the checklist without leaving the page.

<input id="checklist-filter" type="search" placeholder="Filter commands, configs, checks" aria-label="Filter deployment checklist">
<script>
const filter = document.querySelector('#checklist-filter');
const rows = [...document.querySelectorAll('[data-filter-row]')];
filter.addEventListener('input', () => {
  const q = filter.value.trim().toLowerCase();
  rows.forEach(row => {
    row.hidden = q && !row.textContent.toLowerCase().includes(q);
  });
});
</script>

Keyboard Shortcuts

These are pragmatic shortcuts for the operator console, not NVIDIA-defined global shortcuts. Keep them consistent across lab, bench, and field builds.

Commands Grouped By Purpose

Inventory and baseline

• Record OS, kernel, L4T, CUDA, TensorRT, container runtime, model hash, and power profile in one diagnostics bundle.
• Capture the exact carrier board, camera module, storage device, and power supply revision.

cat /etc/nv_tegra_release
uname -a
dpkg -l | grep -E 'nvidia-l4t|cuda|tensorrt'
/usr/src/tensorrt/bin/trtexec --version || true
docker version
docker info | grep -i runtime

Power and thermal validation

Use nvpmodel to select a board-supported mode, then use jetson_clocks only for controlled benchmarking because it changes the normal dynamic behavior you may see in the field.

sudo nvpmodel -q
sudo nvpmodel -m 0
sudo jetson_clocks --show
sudo jetson_clocks
tecrastats --interval 1000 --logfile /tmp/tegrastats.log &
tecrastats --stop

Container deployment

NVIDIA documents docker pull, docker image ls, and docker run as the basic Jetson container workflow; production teams should add digest pinning, health checks, and read-only mounts.

sudo docker pull nvcr.io/nvidia/l4t-base:<tag>
sudo docker image ls
sudo docker run --rm -it --network=host --runtime nvidia \
  --device /dev/video0 \
  -v /run/udev:/run/udev:ro \
  -v /var/log/robot:/var/log/robot \
  your-registry/perception@sha256:<digest>

Model and latency checks

• Benchmark the model on the target Orin module, not on a workstation GPU.
• Track p50, p95, and p99 latency with the real camera rate and robot middleware enabled.
• Store calibration data and preprocessing code beside the deployed engine.

/usr/src/tensorrt/bin/trtexec \
  --onnx=model.onnx \
  --saveEngine=model.plan \
  --fp16 \
  --shapes=input:1x3x640x640

/usr/src/tensorrt/bin/trtexec \
  --loadEngine=model.plan \
  --warmUp=1000 \
  --duration=60

Configuration

Deployment manifest

Keep configuration in a structured manifest and format it before review with the TechBytes Code Formatter. That prevents whitespace-only diffs from hiding changes to model paths, topics, or safety thresholds.

robot: orin-field-unit-17
jetpack: "7.2"
jetson_linux: "39.2"
power_mode: "board-specific-mode-id"
container:
  image: "registry.example.com/perception@sha256:<digest>"
  runtime: "nvidia"
model:
  format: "tensorrt-plan"
  precision: "fp16"
  input_shape: "1x3x640x640"
telemetry:
  tegrastats_interval_ms: 1000
  latency_percentiles: [50, 95, 99]
rollback:
  previous_image: "registry.example.com/perception@sha256:<previous_digest>"

Pre-flight gates

1. Boot from the intended storage path and verify the root filesystem version.
2. Run camera capture, encoder, model inference, and actuator loop at the same time.
3. Confirm no thermal throttling under the longest expected duty cycle.
4. Power-cycle the robot and verify the selected nvpmodel state, services, and containers recover.
5. Pull the network cable and verify the robot fails into the documented local behavior.

Advanced Usage

DLA, GPU, and mixed pipelines

• Use DLA when the model is supported and freeing GPU headroom matters more than absolute flexibility.
• Keep preprocessing and postprocessing in the latency budget; small CPU steps can dominate robotics loops.
• Profile camera ingress, model execution, middleware serialization, and control publication as one pipeline.

Security and privacy hardening

• Mask logs before sharing field bundles; use the TechBytes Data Masking Tool for tokens, customer identifiers, GPS traces, and internal hostnames.
• Use signed images, least-privilege containers, read-only mounts, and explicit device mappings.
• Document how keys are provisioned, rotated, revoked, and audited for each fleet stage.

Sticky ToC pattern

For long deployment runbooks, keep a sticky table of contents in the template and let the live filter handle the dense sections.

<aside class="toc sticky top-6" aria-label="Deployment checklist contents">
  <a href="#version-baseline">Version baseline</a>
  <a href="#keyboard-shortcuts">Keyboard shortcuts</a>
  <a href="#commands-by-purpose">Commands</a>
  <a href="#configuration">Configuration</a>
  <a href="#advanced-usage">Advanced usage</a>
</aside>