The simplest form is behaviour cloning: record many demonstrations as pairs of observation and action, then train a neural network to predict the action a human would have taken. No reward function is needed, which makes it much easier to set up than reinforcement learning for manipulation tasks like folding cloth or plugging in a cable.
Demonstrations come from teleoperation with joysticks or a copy of the arm, from hand-guiding the robot, or from motion-capture gloves. Collecting them is the main cost, and data quality matters more than quantity.
The known weakness is compounding error: once the robot drifts into a situation no demonstration covered, it has no idea what to do. Remedies include collecting corrections in exactly those situations and combining imitation with reinforcement learning.
