Knowledge Driven Robotics at the Georgia Tech Research Institute (GTRI)
Scott Lab E141
201 W. 19th Avenue
Columbus, OH 43210
Seminar speaker: Stephen Balakirsky
Current robotic systems tend to be brittle. They work well in very constrained environments while performing repetitive tasks. However, the systems lack the agility to cope with errors, the flexibility to work with humans or diverse problem sets, and the adaptability to cope with changes in the environment or process. One solution that is often employed is to make the system more complex by adding more parameters that must be tuned by users, or more complex conditions that must be satisfied for mission success. This leads to a need for a highly trained workforce that has both task expertise and robotics expertise. The Georgia Tech Research Institute (GTRI) is creating a planning framework that trains the robot rather than the workforce. It allows the robot to understand both when actions are appropriate, and the expected outcomes and consequences for action execution. Through this understanding, the system is able to plan complex missions, detect failures, and provide corrective actions.
This system is based off of three widely used constructs; the Planning Domain Definition Language (PDDL), behavior trees, and a Mongo world knowledge database. This talk will present our architecture that ties these constructs together, and the planning and control techniques that allow the system to be more agile, flexible, and adaptable than traditional robotic cells. In addition, several use cases and examples of the framework’s application will be presented.