Series 3 Models

Series 3 moved inference to the Jetson Orin Nano, increased image input to 320x180, and introduced the hybrid steering head in v3.1. The table below uses the corrected common evaluation: every checkpoint was run on the same 6,952-frame frozen Series 3/4 challenge subset. These are held-out replay metrics, not physical-car results.

Shared Challenge Metrics

Model Bal9 Turn exact Turn +/-1 ST exact MAE Median AE Signed Field status
3.0 16.7% 16.5% 45.8% 23.5% 18.187 13.122 -2.279 Historical
3.0b 16.2% 15.6% 44.6% 26.3% 17.573 12.629 -2.586 Historical
3.1 28.8% 27.2% 55.5% 55.7% 22.232 9.440 -3.113 Historical
3.1b 28.2% 25.2% 53.8% 57.1% 20.590 9.333 -4.510 Historical
3.2 26.1% 24.3% 55.1% 51.0% 17.461 9.727 -1.454 Previous baseline
3.2b 21.0% 15.4% 43.9% 67.6% 14.640 5.127 -4.539 Previous baseline
3.3 23.8% 18.5% 47.7% 65.3% 15.168 6.130 -5.411 Went farther than 3.3b, but performed worse than 3.2 in the shadow test
3.3b 19.0% 8.6% 36.9% 79.2% 16.078 1.897 -8.194 Left the sidewalk in shadow within about 10 m; much worse than 3.2b
3.4 24.2% 22.6% 56.2% 64.2% 15.083 6.069 +0.418 Field-selected baseline
3.4b 22.4% 19.1% 51.2% 72.7% 13.985 2.478 -1.566 Slightly worse than 3.4 in field

Why v3.4 Won

v3.4b has lower MAE, lower median error, and higher straight recall. v3.4 has stronger turn exact, turn-within-one-bucket, and near-zero signed error. More importantly, v3.4 completed every shadow case presented in the July 13 physical comparison. That combined evidence is more relevant to steering than any isolated column.

v3.3 and v3.3b are useful negative results. They were trained to improve shadow robustness, but the physical car became worse. Stronger augmentation is not automatically better if it obscures geometry needed for real turns.

Model Contract

  • Model file: code/ai_models/SidewalkPilot-v<version>.onnx
  • Input: [batch,3,180,320], normalized OpenCV BGR
  • v3.0/v3.0b output: [batch,2] steering and throttle regression
  • v3.1+ output: [batch,19] = 9 logits + 9 offsets + throttle
  • Runtime default: DEFAULT_STEERING_MODEL_CHOICE = "3.4"
  • Deployment: ONNX Runtime CUDA on Jetson Orin Nano

See the Series 4 table for the temporal experiments and the full PDF report for confusion matrices and all 52 checkpoints.