crossposted from willowgarage.com
This summer, Dan Munoz of Carnegie Mellon University worked on helping the PR2 understand its environment using its 3-D sensors. Improving 3-D perception is important because it can help the PR2 with many tasks such as localization and object grasping. At CMU, Dan and collaborators are developing techniques to improve 3-D perception for an unmanned vehicle in outdoor natural and urban environments. These techniques first take in a cloud of 3-D points, usually collected from a laser scanner, and a label associated with each point. These labels identify such objects as buildings, tree trunks, plants, power lines, and the ground. Then, various local and more global features that describe the local shape and distribution of each object are extracted for each point and region of points. These labeled examples are then used to train an advanced machine learning tool that reasons the best way to combine the local and global features that describe each object. In new environments, this feature extraction process is repeated and given to the machine learning tool to determine what objects are present in the novel scene.
While at Willow Garage, Dan integrated this learning framework into ROS. As shown in the video, Dan experimented with helping the PR2 perceive objects on the room-sized scale, such as tables and chairs, as well as objects at the table-top-sized scale, including mugs and staplers. During the Intern Challenge, Dan also applied this same framework to distinguish between the three different types of bottles being served: Odwalla, Naked, and water. Dan developed the descriptors_3d package, the library used to compute various 3-D features for a point or region of points from a stereo camera or laser scanner. Additionally, he developed the functional_m3n package (Functional Max-Margin Markov Networks), the advanced machine learning tool that learns how to combine low-level and high-level feature information for each object.
Below are Dan's final presentation slides.