Robust and Long-Term Visual Mapping and Localization

This project develops robust and persistent algorithms for mapping and localization using low-cost visual/depth cameras and inertial sensors. New map representations and algorithms are developed to provide computationally efficient long-term 3D mapping and navigation. Topics of investigation include incremental non-Gaussian inference techniques, dense mapping, change detection in dynamic environments, and semantic understanding. A lack of robustness has been a key shortcoming of previous techniques for localization, and thwarted the development of persistently autonomous mobile robot systems. Extension to multimodal distributions poses significant intellectual difficulties. Dense methods are transforming robotic perception, enabling sophisticated physical interaction with objects, traversal of stairs, and safe maneuvering in cluttered and confined spaces. Whereas most past research in robotic mapping has assumed a static world, the approach being developed in this grant exploits the dynamics of the world to discover information about objects and places. These advances are being tested for robotic and man-portable sensing systems operating in indoor, outdoor, and underwater environments. The expected impacts span a broad range of applications, from robotic manufacturing, medical robotics, agriculture, and space and underwater exploration, in which perception is a key requirement. Other potential spin-offs include human-portable mapping applications in real estate, construction, and facility maintenance, health and safety. MIT Online Robotics Education provides a set of online course materials for core topics in robotics, targeted to a broad audience for high school and college education. Open source software modules provide positioning capabilities for low-cost robots for education and service robotics applications.