Deploying Vision AI on NVIDIA Jetson Thor with ROS 2

Introduction

Edge vision pipelines reduce dependency on cloud inference and can lower latency for robots.

Camera
  |
  v
Preprocessing
  |
  v
AI Inference
  |
  v
Detection Results
  |
  v
ROS 2
  |
  v
Planner / Controller

1. Capture camera frames

Use the camera stack supported by your Jetson and robot hardware.

2. Prepare inference

Production Jetson deployments commonly use NVIDIA's accelerated AI software stack, such as CUDA and TensorRT, where appropriate.

Keep model preprocessing identical between development and deployment.

3. Create a detection message

class_id
confidence
x
y
width
height
timestamp

4. Publish through ROS 2

A perception node can publish detections to:

/perception/detections

5. Consume detections

def detection_callback(message):
    for detection in message.detections:
        print(detection.class_id)

6. Optimize the pipeline

Measure:

Camera capture latency.
Preprocessing latency.
Inference latency.
ROS 2 transport latency.
End-to-end control latency.

Do not optimize only model inference; the complete pipeline determines robot responsiveness.

7. Safety

Vision output should not directly trigger dangerous actuator commands without validation. Add confidence thresholds, command limits, watchdogs, and a safe-stop mechanism.

Conclusion

Jetson's accelerated edge computing capabilities can provide a foundation for real-time robot perception when combined with ROS 2 and carefully measured pipelines.