Dealing with uncertainty is a fundamental challenge for building any practical robot platform. In fact, the ability to adapt and react to uncertain scenarios is an essential sign of an intelligent agent. Furthermore, uncertainty can arise from every component of a robotic system. Inaccurate motion models, sensory noises, and even human factors are all common sources of the unexpected. From an algorithmic perspective, handling uncertainty in robotics introduces a new layer of difficulty because the algorithm not only needs to be accurate in a single scenario but also need to adapt to the changes in uncertainties as the environment shifts. This thesis presents methods for adapting to uncertainties in two tasks: object pose estimation and assistive navigation.

For object pose estimation, we present a sensor fusion method that is highly robust in estimating the pose of fiducial tags. The method leverages the different structural and sensory advantages of RGB and Depth sensors to joint-optimize the Perspective-N-Point problem and obtains the pose. The key insight being adaptively bounding the optimization region by testing the pose solution uncertainty.

For assistive navigation, we wish to tackle the problem of using active signaling to avoid pedestrians while it is minimally invasive to other people. We formulate the problem as a bandit with expert advice problem with reinforcement learning policies as the experts. We present an online learning algorithm which can continuously adapt to new and uncertain pedestrian types by using an online policy search technique and the Dirichlet Process

Thesis Committee:
Kris Kitani (Co-Chair)
Siddhartha Srinivasa (Co-Chair)

Frank Pfenning, Head, Computer Science Department
Andrew W. Moore, Dean, School of Computer Science

51 pages

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by [email protected]