Structured Deep Learning for Modeling and Controlling High-Dimensional Dynamical Systems

Abstract

Structured Deep Learning for Modeling and Controlling High-Dimensional Dynamical SystemsWe will leverage our expertise in robotics, nonlinear control theory, geometric mechanics, machine learning, computational cognitive science, and human visuomotor control to develop deep learning architectures that incorporate domain knowledge in the form of (1) physical principles obeyed by the dynamical system (e.g., conservation laws for energy or momentum, and invariances in translation or rotation), and (2) structure endowed by the associated control tasks. To complement this effort, we have proposed a series of carefully crafted experiments aimed at investigating human strategies for exploiting domain knowledge of physical interactions based on perceptual and action-related features of a novel system. The experiments will address fundamentalgaps in our understanding of human visuomotor control of novel objects and will inform development of machine learning strategies.The proposed work targets the fundamental science of data-efficient andgeneralizable control of novel systems based on rich sensory inputs in both engineered autonomous systems and humans. We expect the proposed effort to result in new theoretical and algorithmic frameworks for data-efficient control along with unprecedented demonstrations of vision-based control usingextremely sparse interactions (e.g., a few minutes of data collection) with novel systems including aerial and underwater vehicles, autonomous ground platforms, legged robots, and mobile manipulators. In addition, we anticipate fundamental contributions to our understanding of how humans rely on visual and action-related features of a novel object to ???load??? domain-specific knowledge that supports quick learning. Our tightly integrated multidisciplinary research project will establish new collaborations, strengthen existing ones, and provide a uniquely interdisciplinary training to students and postdocs.The proposed work has the potential to drastically improve DoD capabilities for deploying autonomous agents in unstructured and complex environments wherewarfighters operate; such environments are characterized by changing dynamics, operating conditions, andmissions. Adaptability to changing conditions by exploiting domain knowledge could enable the DoD to deploy mobile robots, including underwater and aerial vehicles, autonomous ground platforms, and legged robots, in settings that were previously impractical. The proposed effort tackles key DoD-relevant challenges and aligns well with ongoing ONR efforts.

Document Details

Document Type
DoD Grant Award
Publication Date
Oct 17, 2018
Source ID
N000141812873

Entities

People

  • Anirudha Majumdar

Organizations

  • Office of Naval Research
  • Trustees of Princeton University
  • United States Navy

Tags

Fields of Study

  • Computer science

Readers

  • Distributed Systems and Data Platform Development
  • Neural Network Machine Learning.
  • Robotics and Automation.

Technology Areas

  • AI & ML
  • AI & ML - Autonomous Systems
  • Autonomy
  • Autonomy - Autonomous System Control