Sanford University researchers are working on developing better ways for autonomous systems goals setting. This project is being spearheaded by Dorsa Sadigh, a computer science, and electrical engineering assistant professor. She and her lab team have combined two different goal setting method for robots into one process. This ends up performing much better compared to when each method is used alone in either real-world or simulation experiments. This work was presented on 24th June at ‘the Robotics: Science and Systems conference.’
With the future only likely to bring more autonomous systems which will require some concepts to help identify good from bad, its, therefore, necessary to get things right as early as now, if at all these autonomous systems will be deployed in future, said a computer science graduate, Any Palan, who is also the paper’s co-lead author.
The reward function, the system used by the team to provide robots with instructions, combines demonstrations- ‘human show the robot what to do,’ and surveys based on the user preference- whereby people answer questionnaires about how they expect the robot to act.
Using both demonstrations and preferences help give more accurate data, which enable the researchers to learn more about the preferred reward function of human. Having a human demonstrate action to an autonomous robot tends to give the robot a considerable chunk of information. However, the robot is not always able to distinguish which part of the demonstration is most important. The surveys, therefore, help to solve this problem.
References
https://www.sciencedaily.com/releases/2019/06/190624124457.htm
Junoh Kim 