Learning new activities (i.e., sequences of actions possibly involving new objects) from single demonstrations is common for humans and would thus be very desirable for future robots as well. However, “one-shot activity learning” is currently still in its infancy and limited to just recording the observed objects and actions of the human demonstrator. In this paper, we introduce a process called “Mental Elaboration and Generalization by Analogy” to create a generalized representation of an activity that has been demonstrated only once. By abstracting over various dimensions of the learned activity, the obtained activity representation is applicable to a much wider range of objects and actions than would otherwise be possible.
@inproceedings{wilsonscheutz14iberamia,
title={Analogical Generalization of Activities from Single Demonstration},
author={Jason R. Wilson and Matthias Scheutz},
year={2014},
booktitle={Advances in Artificial Intelligence},
url={https://hrilab.tufts.edu/publications/wilsonscheutz14iberamia.pdf}
doi={10.1007/978-3-319-12027-0_40}
}