Particle Filter-Based Robot Localization

This project was the first project I implemented for Byron Boots' excellent Statistical Techniques in Robotics class. We were given odometry and laser range finder data self-collected by a small mobile robot moving around a known map, which we were also given, and our task was to find the location of the robot in the map as it moved around. Theoretically we could sample map locations and find the likelihood that a particular observation originated there and from this derive the robot's most likely location.

The data from the robot, however, was of course inaccurate, and required that we model the "fuzzy" odometry dynamics (motion model) and the laser range finder errors (sensor model) as distributions from which we would build the proposal distribution that we would then sample for potential locations on the map for the robot. One of the more difficult components of this assignment was choosing and tuning the hyperparameters used to build the motion and sensor models required. Visualizing the models as distribution curves helped with this, although ideally these parameters would be found using some kind of iterative optimization.