Uncertainty in Robotics: how to understand, how to manage
- Reference number
- SAB25-0019
- Project leader
- Saffiotti, Alessandro
- Start and end dates
- 261201-270831
- Amount granted
- 1 346 750 SEK
- Administrative organization
- Örebro University
- Research area
- Computational Sciences and Applied Mathematics
Summary
This sabbatical will investigate the nature of uncertainty in Artificial Intelligence (AI) systems, the mathematical models to represent it, and the computational models to deal with it. I will focus on AI systems that interact with the physical world, or robots. The objective is to develop an organic treatment of the different types of uncertainty that may affect the functionalities of a robot, coupled with an analysis of the models and computational methods to deal with them. This treatment will be made available to the future workers in physical AI - be them students, researchers, educators or innovators - in the form of a book on "dealing with uncertainty in robotics", something which is missing in the burgeoning field of intelligent robot systems. The ambition of this book is to guide those workers through the choice of the right techniques to use to deal with uncertainty in their specific systems. The study will be done in close collaboration with researchers at the Host institution. I will start by analyzing what uncertainty is present, and how it has been addressed, in a number of projects at the Host institution. This will lead to the definition of a framework to understand and deal with uncertainty in AI systems, which will then be instantiated to the case of physical AI systems. The writing of the book will be done after returning to my home institution, to share my findings with the local robotics researchers, and to use their feednack.
Popular science description
Physical AI systems are AI systems that directly interact with the physical world by sensing and acting: a typical example is an intelligent robot. When an AI system interacts with the environment, it necessarily faces many different types of uncertainty: the camera images may be noisy, the wheels may slip, and generally the robot cannot predict if and when a person will step in front of it. Developers of robots have a number of technologies at their disposal to allow their robot perform well in face of this uncertainty. What they do not have, however, is a guide on what technology to use in what case, depending on the specific type of uncertainty: for instance, the best tool to deal with a noisy image may be different from the one to deal with unforseen human motions. This study will seek a fundamental understanding of the types of uncertainty that a robot may be confronted with, and of the best tools to use to deal with each such type. The tangible output of this study will be a book on "dealing with uncertainty in robotics", addressing the future workers in physical AI --- be them students, researchers, educators or innovators. The ambition of this book is to provide those workers with the ability to understand the nature of uncertainty in their specific system, and to guide them in the choice of the right techniques to use to deal with their specific type(s) of uncertainty.