CAREER: Model Probability Planning for Mobile Robots

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

Abstract for Proposal # 0546467 Model-Probability Planning for Mobile Robots Robots operating in natural, dynamic domains depend on the ability to make decisions without perfect prior knowledge of the world. Robots instead use sensor data to infer a model of the world (e.g., a map) and then make decisions with respect to this map. Typically, the process of learning the map is separated from planning, simplifying both problems. In natural, populated environments, however, building a complete and accurate map becomes increasingly difficult and motion planning can become brittle. For example, a robot grasping an object on a table should not depend on having an exact and precise geometric model of the table and object. In contrast, the robot should be able to learn the model as it goes, starting with a simple probabilistic model of the table top and object. Then, to reduce uncertainty, the robot should gather new data about the world, stopping only to take more precise measurements of the world to improve its map when doing so is predicted to lead to better overall performance. The objective of the proposed research program is to develop robust autonomy in robotics by integrating mapping and planning in a tractable manner. The approach will be to develop planning systems that operate in the space of model distributions. New techniques of modeling free and occupied space probabilistically must be developed to capture map uncertainty efficiently. Unlike conventional maps that predict sensor error, the resulting maps will be designed to maximize the predicted planner performance. Secondly, motion planning algorithms must be developed that can plan with respect to a range of possible maps, generating motion plans that incorporate exploration and resolve ambiguities. The proposed scientific activity will create a new generation of more robust and capable autonomous systems through a combined research and educational program to study integrated learning and decision making. The educational program will introduce a generation of students to new principles of planning under uncertainty and mobile manipulation. The proposed research is vital to the long-term viability of mobile robots operating autonomously in complex, dynamic environments. For example, mobile manipulators have great potential in a number of assistive domains such as health care and manufacturing, but reliable operation in populated environments will require a solid understanding of decision making under uncertainty through integrated learning and planning. This research will also lead to robust autonomy in unmanned underwater, air and space vehicles.