Conformational Probabilities: A Bridge Between Innate Knowledge and Action Recognition
Abstract
The goal of this proposal is to develop a methodology that blends elements from both classical AI/expert systems and modern probabilistic machine learning to endow robots with intelligent behaviors that are robust and grounded (physically and semantically).Robotics differs substantially from other areas of intelligent systems research such as computer vision and natural language processing. Deep learning has made tremendous progress in recent years in these other areas, but a case is articulated here as towhy a hybrid learning-AI approach is more suitable for robotics. In short, robots are automatically grounded in the physical w rld, a robot can take actions that have ramifications to both its future physical state and state of knowledge, and the types of sensing modalities available in robotics are substantially different than in other areas of machine intelligence research. The proposed approach involves defining classes of humanoid (human and anthropomorphic robot) actions expressed as conformational probabilities and identifying such actions through a mathematicalformalism called quotient operations. The result of applying quotient operations is a ~quotient space~ (also referred to in some literature as a ~factor space~). This recognizes that individual human and robot actions such as locomotion and manipulation tasks can have infinite variations in their spatiotemporal trajectories. The proposal therefore seeks to do the following: (1) develop a methodology for partitioning ensembles of trajectories associated with actions into subsets containing trajectories that are equivalent in terms of their net effect (or meaning); (2) represent classes of actionsprobabilistically as distributions on conformation space, and use the resulting information to recognize tasks that are currently difficult for machine intelligence systems to parse by using quotient operations; (3) demonstrate themethodology in a hardware testbed. This basic research is motivated by the tremendous impact that future deployment of human-robot cooperative teams on Navy ships would have
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2017
- Source ID
- N000141712142
Entities
People
- Gregory Chirikjian
Organizations
- Johns Hopkins University
- Office of Naval Research
- United States Navy