Imitation learning is a process wherein a neural network is trained to predict what a human driver would do in the current situation (and then it does that). Imitation is behaviour prediction.
Behaviour prediction as it applies to vehicles (such as cut-in prediction) also predicts what a human driver will do in the current situation.
The differences are:
Behaviour prediction uses purely visual information, whereas imitation learning can use driver input (steering, braking, accelerating, signalling).
With behaviour prediction, the goal is accuracy, whereas with imitation learning the goal is quality. So, for imitation learning (unlike behaviour prediction) you might only use data from the best drivers.
An imitation network’s predictions are used to drive, whereas a behaviour prediction network’s predictions are used to anticipate the driving of others (and respond accordingly).
The idea that you’re going to build a full predictive model of road users in order to enable a heuristics-based path planning/driving policy system to work is questionable because once you can build a full predictive model, you can also presumably build a full imitation model that can handle path planning/driving policy.