[Blue Sky] Are robots stuck in time? Learning to represent actions in long horizon planning by learning to remember the past
Keywords: Long horizon tasks, episodic memory, representation learning
TL;DR: In order to do long horizon planning robots should improve their representations of long action episodes by learning to remember their past actions.
Abstract: In order to do effective long horizon planning it will be necessary for robots to use representations of action sequences of various durations and levels of abstraction. Many of these will need to be of quite long duration, representing entire episodes or sequences of episodes of action. Like the most successful image and text representations, these will need to be learned. Humans learn and use such representations of action episodes partly in taking actions but also in remembering their past actions. If robots are to develop the kind of expressive and flexible representations of action that humans have, they will also have to learn to represent and remember their past actions. These representations of past actions will, in turn, be able to serve as the basis for useful kinds of reasoning about the future, as they do in humans.