CSAIL Event Calendar: Previous Series
Mines, Maps, and Log-Likelihood
Speaker: Sebastian Thrun , CMU
Hundreds of thousands of abandoned mines exist in the US, most of which lack accurate maps. Most of these meins are inaccessible to people, which motivates the development of robotic technology for mine exploration and mapping. From a mathematical perspective, the problem of robotic mapping of subterranean spaces is an instance of the simultaneous localization and mapping problem, or SLAM. In this talk, I will discuss a SLAM algorithm that represents a robot's map in log-likelihood form. Our approach approximates a Bayesian posterior over a robot's map using a sparse Gaussian Markov random field. Updating and inference is efficient in this framework, as attested by a number of recent theoretical results. The Markov random field representation, which is reminiscent of spring-mass models studied in other fields for decades, give rise to new lazy data association techniques in robotics. Our approach has enabled robots to acquire detailed 3D maps of abandoned mines. It have also been brought to bear in related aerial mapping problems using robotic helicopters.