Hybrid Head Versus Regression

Series 1/2 and v3.0 directly regress steering. Series 3 v3.1 and later use a nine-class steering distribution plus one local offset per class.

Contracts

Series 1/2 produce one bounded steering value. The v3.1+ Series 3 head produces 19 values:

9 class logits + 9 class-local offset logits + 1 throttle logit

The steering classes are:

0-45 | 45-60 | 60-75 | 75-85 | 85-95 | 95-105 | 105-120 | 120-135 | 135-180

At inference, the decoder applies softmax to the nine class logits and selects argmax(softmax(class_logits)). It then maps the selected class's sigmoid-bounded offset within that class. Softmax preserves the argmax ordering, but writing the complete operation makes the probability interpretation clear. The current runtime ignores the throttle output; motion policy owns throttle and braking.

Engineering Reason

A direct regression loss can favor values near a common central target on an imbalanced dataset. A hybrid head makes coarse steering-class behavior visible in the loss and in evaluation while retaining a continuous within-class angle. It also introduces an argmax class boundary: a small probability change can switch the selected class. The architecture therefore changes the error surface; it does not guarantee turn recall or smooth closed-loop steering.

Evidence

  • v3.0/v3.0b use the regression contract; v3.1 and later use the hybrid contract.
  • The common evaluator measures the hybrid models with nine-class recall, adjacent-class recall, continuous error, and signed bias.
  • v3.3 and v3.3b regressed in the July 13 field comparison despite using the hybrid head.
  • v3.4 became the field-selected model after completing the presented turn and shadow cases.

These observations support keeping class-level evaluation. They do not isolate the head design as the cause of any checkpoint's field result because data, training settings, and weights also changed across versions.

Alternatives

Head Advantage Limitation
Direct regression Small and simple Aggregate losses can hide rare-class behavior
Pure classification Direct class supervision Quantized output unless a continuous stage is added
Class plus local offset Class metrics plus continuous angle More complex loss/decoder and class-boundary discontinuities