Abstract:

Robust robot operation must answer: What to do now, to receive good long-term returns, despite notRobust robot operation must answer: What to do now, to receive good long-term returns, despite notknowing the exact effect of its actions, despite various errors in sensors and sensing, and despitelimited information about the environment and itself. This problem is not new. Mathematically principledconcepts --called Partially Observable Markov Decision Processes (POMDPs)-- have been developedmore than five decades ago to address the problem mentioned above. However, such concepts arenotorious for their computational complexity, that they have often been considered impractical. I willpresent some of our effort in addressing the computational complexity issues of solving POMDPs, anddemonstrate that this decision making concept has now become practical (to some extent) for solvingvarious problems in robotics. I will end with a discussion on what this technology could mean inbridging the gap between sensing and acting in robotics, and between planning and learning ingeneral.