Model Framing and Loss

All model families use the same logical steering convention: 0 degrees is left, 90 is center, and 180 is right. Their output heads and losses differ.

Direct Regression

Series 1/2 and v3.0 predict continuous controls directly and use Smooth L1-based regression. Direct regression is compact, but common near-straight labels can dominate its aggregate objective.

Class Plus Local Regression

Most Series 3 models predict 19 values: nine steering-class logits, nine local offsets, and throttle. The decoded result remains a continuous steering angle. Its training objective combines focal-weighted class loss, Smooth L1 loss for the true class's offset, and an optional throttle loss:

total = focal_cross_entropy(class)
      + offset_weight * smooth_l1(selected_offset)
      + throttle_weight * smooth_l1(throttle, target_throttle)

Series 3 defaults include class-weight power 0.3, focal gamma 1.5, and offset weight 1.0. Steering-focused runs explicitly set throttle weight to zero; current defaults are not evidence of an older checkpoint's command.

Series 4 Temporal Framing

Series 4 removes throttle and uses an 18-value class-plus-offset steering head per horizon:

• PC (4.0p/r) supplies the image plus causal previous steering targets and predicts the current target;
• CF (4.0f/g) supplies the image and predicts current plus future targets;
• PCF (4.0a/c) combines causal previous-target inputs with current and future supervision.

Future steering values are labels during training. They are never future inputs at deployment. Fixed Series 4 runs use class-weight power 0.5, focal gamma 1.5, no sampler balancing, and future-horizon decay 0.70 where applicable.

Gradient Norm and Limits

The trainers log gradient norm before clipping and clip to a maximum norm of 1.0. Gradient norm diagnoses update stability; it is not a quality score. The losses also do not encode stopping distance, sidewalk boundaries, or physical smoothness. Offline class-balanced evaluation and supervised driving remain required.

See CNN Architecture, Series 3 Hybrid Head, and Series 4 Temporal Experiments.