Learning from demonstration (LFD) is an approach to training robots. In learning from demonstration a human demonstrates a specific task to a robot that attempts to learn to imitate the human?s actions. This allows humans to train robots without requiring the human to have any specialized knowledge beyond knowledge of the task domain, e.g. they don?t have to know any programming. However, complex tasks or tasks performed in complex and changing environments require either a very large number of demonstrations (to capture the complexity of the task and the range of possible environments) or a learning algorithm that is very good at generalizing from a relatively small number of demonstrations. Evolutionary computation?s ability to generalize from a small training set, i.e. a small number of demonstrations, makes it a very promising approach for learning from demonstration. But evolutionary computation requires a computationally powerful robot to run the learning algorithms. One approach to this problem is the use of Commodity-off-the-shelf robots (COTSBots). COTSBots are relatively low-cost robots, built from inexpensive, commodity parts, that have the computational power to run evolutionary algorithms on-board and in real-time. This presentation covers both how to build COTSBots and research directions in evolutionary learning from demonstration.


Terence Soule

Evolutionary computation methods e-session