Model Deployment Overview

The current heavy-model deployment path is:

PyTorch training checkpoint
  -> FP32 ONNX export
  -> copy ONNX to Jetson Orin Nano
  -> ONNX Runtime with CUDA
  -> decoded steering returned to Raspberry Pi 5

The Raspberry Pi 5 remains responsible for the camera, steering servo and motor control, result freshness, manual override, and LiDAR safety. The Jetson Orin Nano runs the steering models and returns their predictions.

Model Contracts

Family	ONNX input	ONNX output
Series 3 v3.0	image	[batch,2]
Series 3 v3.1+	image	[batch,19]
Series 4 PC	image + target_history[batch,3]	[batch,1,18]
Series 4 CF	image	[batch,4,18]
Series 4 PCF	image + target_history[batch,3]	[batch,4,18]

The server inspects input names and output shapes rather than assigning a contract from the filename alone.

GPU Selection

Series 1/2 run through PyTorch CUDA. Series 3/4 run through ONNX Runtime's CUDA provider. CPU execution remains available only for diagnosis; a field launch should confirm that the Jetson Orin Nano GPU path loaded.

Export and Verification

Before field use, each model must load with the expected input names and output shape, return finite steering values, and pass a complete Raspberry Pi 5-to-Jetson Orin Nano response test. The deployed filename and version must match. The current field path uses FP32 models; TensorRT, FP16, and INT8 are not active.

See Jetson Orin Nano Runtime and Jetson Orin Nano Inference Link.