Robot Perception

Our approach in the field of Robot Perception encompasses numerous research goals and pursuits. For instance, we investigate lifelong mapping and environment reconstruction. Current robot mapping systems take one sweep through an environment, construct a map off-line, and then use it until the environment changes enough to require another sweep. Our goal is to continuously update a detailed geometrical map as the robot goes about its tasks, using both vision and laser sensors to recognize places, and label spaces and objects therein.

Another perception focus is reliable object detection, registration, and tracking. Manipulating tabletop objects is one of the main goals of the PR2 robot. Our aim is to have near-perfect, real-time recognition of common small objects. One critical part of this project is the creation of CAD models, either through an off-line modeling system, or with "in-hand" creation for novel objects.

Additionally, we work on detecting people and understanding their actions. Our robots are meant to co-exist with people in everyday environments. Initially, we detect people to navigate in a more friendly fashion; we are working towards tracking face pose and body motion as part of interacting with people in cooperative tasks.

Relevant Blog Posts

• Person-following and Detection in an Indoor Environment
• Hand Detection and Image Descriptors
• Detecting Tabletop Objects
• Indoor Object Detection
• 3D Object Recognition
• Collision Detection and Arm Planning
• Detecting Planar Objects and Articulation Models

Publications