To be a useful household assistant or to aid in disaster recovery, robots often need to search for occluded objects by moving obstacles out of the way. We are interested in enabling robots to reason about the uncertainties in the world (e.g. noise from sensors, the unknown location of the hidden object), and choose actions to find the target object with minimum cost. This problem is often formulated as a Partially Observable Markov Decision Problem (POMDP). However, the state-of-the-art solvers of general POMDP problems are unable to solve realistic instances of the above search problem. Our goal is to develop an efficient algorithm that approximates the optimal solution, and demonstrate it both in simulation and on a PR2 robot.

Since sophomore year, I have worked in the Learning and Intelligent Systems Group as a UROP on projects involving the PR2 robot and motion planning algorithms. I have previously interned at eBay working on topic modeling in recommendation engines, and most recently at Akamai, researching and architecting a system that interfaces the Hadoop Distributed File System and an external object storage system.