Projects

Programming Robots by Description and Learning

Generic controllers (such as a PID controller) allow for a simple way to provide control to a motor quickly, but the performance is not optimal. Learning can provide optimal control, but requires large quantities of training data.

We developed a system which takes a task description in natural language (such as "Follow the red ball.") and turns it into a running controller on a real robot. This system can then be used to train a Reinforcement Learning algorithm to improve performance. Check out the website for the latest information.

Localization and Mapping with Environmental Features

The problem of localizing a robot in an environment is difficult without having a map of the area already. When the task is to construct a map, using dead-reckoning or other similar methods can lead to large cumulative errors over time. Using localization in the map under construction can lead to compounded errors, since even small errors in position can be magnified when used as the basis for creating a map.

We developed a framework for recognizing higher-level objects which can then be used to predict the change in a mobile robot's pose. The positions of the higher-level objects impose constraints on the new position of the robot, and allow us to perform more robust localization than raw sensor readings. This framework has proved fairly robust over the data we have collected on our robot platform. Check out the website for the latest information.

Robot Mapping Software

Making maps from sensor reading on mobile robots is basically a solved problem. The only real problem is the lack of ready-made software to accomplish this task. Our goal is to write robust map making and localizing software that can be used on the research robots in use today to speed up the learning curve on new robots, and eliminate the need for writing volumes of code before the robot moves. The documentation for the current version Robot Mapping Library is available. Currently, only occupancy grid maps are supported, but we hope to add our current research to the library, which would add the ability to use object-based maps.