SIGNAL
Averaged-loss training produces fluent drivers and careless ones. The rare rule-bound instant barely registers in the loss, so the policy never learns to take it literally.
A red light asks a car to commit to stillness on demand; averaged trajectory error barely notices the ask, which is why behavior-cloned models learn to drift through the stop line on the turns a regulator cares about.
RedLight-VLA's fix has a shape, even when the paper gives it a name: weight the rare rule-bound maneuver in the training loss, and ground the rule in an auxiliary head that does not share the trajectory decoder's vocabulary. Do both, and the stop-line overshoot falls. Do only the auxiliary grounding, and the car starts hesitating at greens. Combining the two keeps the gain while shrinking the hesitation cost.
The mechanism travels. Any behavior-cloned system graded on a dense average will underweight sparse rule-bound instants and overweight routine fluency. The transferable move is splitting the supervision: averaged loss for fluency, a separate grounding loss for the rule. The remaining question is whether the hesitation tax is acceptable in the scenarios a fleet actually drives.
Reported by Samantha for Type0, from RedLight-VLA: Models for traffic-rule grounding and behavioral emphasis in driving policies. Read the original: arxiv.org