THIS GRANT IS A CONTINUATION OF N000141410120 Mental Simulation of Intentions for Collaborative Human-Robot Learning and Planning
Abstract
This work is grounded in the idea that humans use motor resonance, simulating a collaborator with ones own motor cortex during joint action. The core concept is a common representation of tasks and common way of doing them. To achieve this common ground, we propose having the robot learn tasks from the human on which it will be expected later to collaborate. The PI s chosen representation will let the robot utilize dynamic mental simulation to create plans for its own actions. The robot will use its model of self to understand the task the human is performing, anticipate thehuman?s intentions and actions and assist in the most useful manner possible.In contrast to traditional imitation learning methods, we will intertwine machine learning and planning/control by means of intention imitation which will be represented as task constraints and heuristics. The link between learning and action will occur at a higher level of inference where mental modeling of geometry, kinematics, dynamics and goals are used to guide robot planning of all joint motions rather than generalization of pre-defined motion primitives.All proposed work will be implemented and evaluated on bimanual, anthropomorphic robots. The PIs? labs operate four human-scale robots, Simon, Mobile Simon, Golem Krang and Golem Hubo which is a whole-body humanoid robot. This will not only ensure that our methods will be generally applicable, but also allow sufficient flexibility in tasks from simplified collaboration when moving lighter objects to complex whole-body motion and heavy lifting.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 23, 2016
- Source ID
- N000141612785
Entities
People
- Mark Riedl
Organizations
- Georgia Tech Research Corporation
- Office of Naval Research
- United States Navy