Learning What to Sense and Communicate for Multi Vehicle Teams
Abstract
A foundational component of multi vehicle networks is the Perception - Action - Communication (PAC) loop, which allows robots to use information gained by communications with their neighbors to decide on future actions towards completing a high-level task. PAC loops extend the concept of operations enabled by OODA loops, a familiar paradigm in military operations. However, the design and realization of PAC loops introduces many challenges. As robots take actions, their relationships with their neighbors and environment change and, consequently, the information they receive also changes. These dynamics are dicult to characterize, and few systematic approaches exist for designing PAC loops except in the simplest of settings. Further, PAC loops may need to be executed over noisy or resource-constrained communication channels, with the time required to communicate being substantially larger than the time scale of the robot dynamics.In this project we propose to develop a general framework for the design of control policies for large scale PAC loops in multiagent systems. Our priority is to propose a design methodology that can more closely approximate real collaborative systems including the ability to handle complex dy-namics, complex perception, realistic communication constraints, and sensible safety requirements. To achieve this goal the proposed framework will be built around three fundamental innovations:(I1) The reformulation of problem of designing decentralized control policies to a problem of deep learning.(I2) The exploitation of centralized policies that we use to learn decentralized policies using a framework of self-supervised or imitation learning.(I3) The use of aggregation Graph Neural Networks (GNNs) as a key technical innovation that enables learning for networked dynamical systems where the information structure can be modeled by an adjacency matrix.Innovations (I1)-(I3) will be leveraged to design control policies and communication policies, as well as to handle complex perception, dynamics and safety requirements in settings that may include constraints on communication, disparate time scales for communications and control, and noisy communication links, and may require visual observations to augment or replace communication links that are denied or disrupted.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 08, 2020
- Source ID
- N000142012372
Entities
People
- R.vijay Kumar
Organizations
- Office of Naval Research
- United States Navy
- University of Pennsylvania