Seminar on Computational Learning and Adaptation


  Towards Learning Object Models of Physical Spaces with Mobile Robots

Sebastian Thrun
Carnegie Mellon University
University of Pittsburgh
Stanford University
thrun@Stanford.edu

While much of machine learning takes place in virtual spaces, learning models of physical spaces is a fascinating and highly active research area. Recent research on robotic map learning has led to a flurry of statistical algorithms for inferring models from sensor measurements. However, the vast majority of work learns models of static environments, at the pixel level. The work presented in this talk seeks to acquire object models, where objects might be movable items (chairs, doors, people?), or structural primitives (walls). The tool of choice for learning such object models is the expectation maximization algorithm, capable of learning model identity, pose and appearance parameters with latent data association problems. Come and enjoy an introduction into a body of work on environment modeling, along with videos of roaming robots.

This is joint work with Dragomir Anguelov, Rahul Biswas, Daphne Koller, Benson Limketkai, Christian Martin, Michael Montemerlo, and Scott Sanner.



Date: Thursday, March 14

Time: 4:15-5:30PM

Place: Cordura 100


Return to the seminar schedule