What is imitation learning in robotics?

The ACT paper describes imitation learning from real demonstrations collected through a teleoperation interface. It learns action sequences for manipulation. This is an example of a specific collection and learning method, not evidence that every demonstration dataset produces the same result.

Keep training examples and test attempts separate. Ask which tasks, objects and locations occur in both. A clean train/test split should match the generalization claim being made. If a report claims new-object performance, check that the object was actually held out. If it claims a new environment, ask which parts of that environment differ. A random frame split can put closely related observations on both sides.

Inspect the observation/action alignment, control mode, timestamps, failed attempts and quality checks. A large video archive is not automatically an action-labelled robot dataset. Human actions captured on video may need further labels or learning methods before they can support a particular robot policy. Dataset suitability depends on the intended task and learning workflow; quantity alone cannot settle it.