Model Zoo

SidewalkPilot has 52 evaluated checkpoints across four series. The live selector currently contains 46 versions through Series 4.0; the six Series 4.1 models are trained and offline-evaluated but not yet integrated. Version numbers identify experiments, not guaranteed improvements.

Families

Series Image input Output design Parameters Current role
1 200x66 direct steering regression 672,877 original camera-steering proof
2 200x66 refined direct regression; CLAHE in 2.0/2.0b 672,877 historical comparison family
3.0 320x180 steering/throttle regression about 5.53M first large Jetson Orin Nano architecture
3.1-3.4 320x180 9 classes + 9 offsets + throttle 5,534,115 field-tested family; v3.4 selected
4.0 PC 320x180 + 3 targets one 18-value steering horizon 5,569,186 rejected after steering echo in field testing
4.0 CF 320x180 four 18-value steering horizons 5,537,560 4.0f viable; complementary to v3.4
4.0 PCF 320x180 + 3 targets four 18-value steering horizons 5,572,696 rejected after steering echo in field testing
4.1 PC 320x180 + 3 targets corrected history training; one horizon 5,537,460 trained and offline-evaluated; integration pending
4.1 CF 320x180 corrected future-target loss; four horizons 5,537,560 trained and offline-evaluated; integration pending
4.1 PCF 320x180 + 3 targets corrected history/future training; four horizons 5,544,480 trained and offline-evaluated; integration pending

Field and Research State

Regular v3.4 is the current field-selected model. It won the July 13, 2026 comparison across every shadow case presented and the tested normal left and right turns.

Series 4.0 was field-tested. 4.0f was viable and traded two passed/failed cases with v3.4, but it did not establish a clear promotion case. 4.0p/r/a/c repeated prior steering predictions and were rejected. Series 4.1 was trained to address that failure and still needs runtime integration and physical testing.

Paired Checkpoints

Series 1-3 use a numeric model and a b checkpoint. Series 4 uses contract-specific letters:

Run Final epoch Validation-selected
4.0 PC 4.0p 4.0r
4.0 CF 4.0f 4.0g
4.0 PCF 4.0a 4.0c
4.1 PC 4.1p 4.1r
4.1 CF 4.1f 4.1g
4.1 PCF 4.1a 4.1c

For Series 3 and 4.0, validation selection primarily follows steering or current-target MAE. Series 4.1 PC/PCF also include closed-loop rollout error in checkpoint selection. Series 1/2 used validation loss. No rule makes the alternate model field-best.

Regular model files preserve the final requested epoch. Their paired b/r/g/c models preserve the epoch selected by that trainer's validation objective. Both are retained so validation behavior can be compared with field behavior. The suffix records provenance, not superiority. A release must keep version, filename, model card, architecture contract, and model hash aligned; the field-selected default changes only after a recorded drive.

Selection Evidence

Aggregate MAE is secondary because the challenge set is straight-heavy. Review Bal9, turn exact, turn within one class, straight exact, signed error, confusion behavior, and a fixed field route together.