Series Differences
Series 1/2, Series 3, and experimental Series 4 are different answers to the same question: look at a sidewalk and decide how to steer. They differ in architecture, input history, output horizon, deployment target, and evaluation criteria.
How It Works
Series 1/2 (SteeringAutonomyV2) |
Series 3 (SidewalkPilotV3) |
Series 4 experimental (SidewalkPilotV4) |
|
|---|---|---|---|
| Params | ~0.67M | ~5.5M | ~5.54-5.57M, contract-dependent |
| Image input | 200x66 | 320x180 | 320x180 |
| Temporal input | none | none | none for CF; three previous targets for PC/PCF |
| Head | single tanh regression |
9 logits + 9 offsets + throttle | 18 values per steering horizon, no throttle |
| Horizons | current steering | current steering/throttle | current only for PC; current + three future for CF/PCF |
| Runs on | Jetson Orin Nano | Jetson Orin Nano | Jetson Orin Nano for v4.0; v4.1 integration pending |
| Throttle | fixed runtime value | present in contract, disabled in steering-focused training | runtime-owned; removed from learned output |
The nine Series-3 buckets are HL, L, L+, SL, ST, SR, R, R+, HR. Series 1 uses an output
scale of 86.0°, Series 2 uses 85.0° (SERIES_1/2_STEERING_OUTPUT_SCALE_DEG), and the
runtime picks Series by the model-choice prefix (steering_model_series(): 2. → Series 2,
else Series 1).
Design Progression
Series 1/2 established the compact regression path. v3.0 tested a larger regression model on 320x180 input. v3.1 and later changed the steering contract to a class plus within-class offset, which makes class recall directly measurable while retaining a continuous angle. The Series 3 graph also contains a throttle output, but current training and deployment do not use learned throttle. The current unified server runs every model family on the Jetson Orin Nano GPU: PyTorch CUDA for Series 1/2 and ONNX Runtime CUDA for Series 3/4.
Key Findings
- Turn-vs-shadow observation. Some field-tested iterations that reacted more strongly to turn cues also followed shadow edges; more center-biased checkpoints could miss turns. Targeted turn-in-shadow data is the current collection response, not a claim that one data type is mathematically guaranteed to solve every case.
- MAE is insufficient. A center-biased checkpoint can score well on a straight-heavy set while retaining weak turn recall. Series 3/4 comparisons therefore include Bal9, turn exact, turn +/-1, ST exact, signed error, and physical testing.
v3.4 is the current field-selected baseline. In the July 13 comparison it handled every presented shadow case; v3.4b was slightly worse, v3.3 was worse than v3.2, and v3.3b was much worse than v3.2b. All six v4.0 models were later driven: v4.0f was viable but mixed against v3.4, v4.0g was worse, and the PC/PCF models echoed previous predictions. Six v4.1 correction models are trained and evaluated offline, but they are not yet integrated or driven. The Jetson Orin Nano runtime currently supports all three v4.0 contracts.