How Imitation Learning Copies Human Drivers

A junior cook stands next to a better cook and copies timing, heat, and motion. Nobody hands them a full theory of sauces. Imitation learning is that apprenticeship at fleet scale.

Clips of people driving — filtered toward the ones a team considers competent — become targets. The network is rewarded for producing similar steering and speed in similar camera views. Do this across enough scenes and you get a policy that looks surprisingly human in ordinary traffic.

Nobody hands them a full theory of sauces.

The catch is that humans are inconsistent teachers. Some hesitate. Some floor it. Some roll through a stop that another driver would fully halt. If you imitate all of them, you imitate the argument. If you imitate only the smoothest drivers, you may miss the assertive gap-taking that actually makes an unprotected left possible.

Imitation also copies what is easy to see and underweights what a good driver only thought. A person covers a brake because they imagined a hidden driveway. The video shows a foot. It does not show the imagination. Later posts on bad data and hidden intent start here.