Jetson Orin Nano Runtime
The Jetson Orin Nano is the inference computer. It is linked directly to the Raspberry Pi 5 over Ethernet at 10.42.0.2:8770; the link does not depend on Wi-Fi.
Live Data Path
- The Raspberry Pi 5 captures the current camera frame.
AsyncJetsonSteeringClientkeeps only the newest pending frame and selected model version.- Jetson Orin Nano's
jetson_inference_server.pyloads the selected model, preprocesses the frame, and runs it on the GPU. - Jetson Orin Nano uses the decoder required by that model's output shape and returns steering plus model telemetry.
- The Raspberry Pi 5 accepts only a recent response for the selected version, applies its safety rules, and then commands the steering servo and motors.
Supported Families
- Series 1/2: one steering output,
200x66input. - Series 3.0/3.0b: two-output regression,
320x180input. - Series 3.1 through 3.4b: 19-output hybrid head.
- Series 4 PC, CF, and PCF: temporal/horizon hybrid contracts, decoded at horizon zero for live steering.
All listed families are supported on the Jetson Orin Nano GPU. Series 1/2 use PyTorch with CUDA; Series 3/4 use ONNX Runtime with CUDA. CPU remains a diagnostic fallback. TensorRT is not the current live path.
Failure Behavior
The Raspberry Pi 5 client does not block the control loop while waiting for Jetson Orin Nano. If no fresh response exists, autonomous driving does not receive a valid new steering command. Manual control and local Raspberry Pi 5 LiDAR braking remain on the Raspberry Pi 5.
Model selection is performed from the dashboard or RC_CAR_STEERING_MODEL; the live runtime currently defaults to v3.4.