This talk will discuss the problem of multi‐robot active estimation, in which sensing robots must control their motion in order to gather the most valuable information about a time‐changing environment. I will cast this as a stochastic optimal control problem where the robots' objective function is to minimize the expected entropy of an environment state estimate formed from a recursive Bayesian filter. I will discuss the prohibitive complexity of solving this problem in general, and will consider two solutions for special cases. The first is a myopic heuristic in which robots follow the gradient of mutual information. The second is to consider periodic trajectories which, in the case of a linear‐Gaussian environment model, can be arbitrarily close to the optimal solution. I will present an iterative randomized trajectory planner which finds periodic trajectories whose cost monotonically decreases in the number of iterations. Efforts toward overcoming the challenges of decentralization and scalability will also be discussed.
About the Speaker
Mac Schwager is an assistant professor in the Department of Mechanical Engineering at Boston University. Heobtained a PhD degree from MIT in 2009, an MS degree from MIT in 2005, and a BS degree in 2000 from StanfordUniversity. He was a postdoctoral researcher in the General Robotics, Automation, Sensing, and Perception (GRASP) Labat the University of Pennsylvania from 2010 to 2011. His research interests are in distributed algorithms for control, estimation, and learning in groups of robots and animals.
Host: Professor Naira Hovakimyan
This seminar counts towards the requirement for ME 590 and TAM 500.