Behavioral cloning

You want to teach a robot to avoid obstacles, but finding behaviors from scratch would be expensive. You have records of what a good controller did in different situations. You can begin by imitating those examples.

Behavioral cloning trains a model to map “observation → action” from demonstrations. This stage does not require discovering rewards independently: the correct action is given in the data.

If the controller turned left for a particular arrangement of pedestrians, the model learns to predict that command or sequence. In the PDPO study, §3, demonstrations prepare the starting point before further learning from rewards.

Imitation depends on the demonstrations' quality and scope. When the robot encounters a situation the data did not cover, it may make errors leading to further unusual states. Good performance on recorded examples therefore does not itself guarantee safe operation in traffic.

See also: Reinforcement Learning — learning from assessments of actions' consequences, for which demonstrations can prepare a starting point.