Decision-making under Uncertainty in Robotics

Uncertainty is ubiquitous. A robot or agent must decide what it should do now to accomplish its tasks, despite not knowing the exact effects of its actions, errors in sensors and sensing, and the lack of information and understanding about itself and its environment. However, the technology for making good decisions in the presence of uncertainty is still lacking. I will present lectures on decision-making under uncertainty. We’ll briefly cover decision-making when the effects of actions is uncertain framework, the Markov Decision Processes (MDPs), progress to its expansion when the states are also only partially observable, namely the Partially Observable Markov Decision Processes (POMDPs). We will end with Reinforcement Learning. In these lectures, the emphasis will be on computational complexity of these robust and rigorous frameworks, and the computational representations and methods that enable the above frameworks to become computationally tractable, at least for many realistic robotics problems.

Requirements and pre-requisites:

Basic probability, Introduction to abstract data structures and algorithms.

Relevance

The lectures will be relevant to:

• Mathematicians interested in bringing rigorous theoretical framework for decision-making under uncertainty to real world applications in robotics
• Practitioners who wants to improve the robustness of their systems in a principled manner.
• Those interested to understand how autonomous agents can handle uncertainty —which is ubiquitous in our world