Joint Perception and Temporal Logic Planning for Distributed Agents in Dynamic Environments
Abstract
The objective of this proposal is to develop theory and algorithms for decentralized perception and temporal logic planning for a team of autonomous agents collaboratively operating in dynamic, uncertain and unstructured environments where large volumes of heterogeneous streaming data are collected. It addresses---through a unified theoretical framework ofdecentralized perception and planning---a key challenge in developing autonomous systems: how to derive intelligence from massive, distributed, and diverse data sources, and enable rapid decision-making. The project has three research thrusts that are tightly integrated into a rigorous closed-loopframework. Thrust I develops algorithms for decentralized perception to derive mission-relevant intelligence from large volumes of heterogeneous streaming data. The key innovation includes new algorithms to integrate information from multiple sensing modalities and diverse sources, filter out irrelevant information, and learn distributed and shared representations of the environment in a team of agents. Thrust II is on planning for the distributed team of autonomousagents, based on the semantic world model with uncertainty learned in Thrust I, subject to a rich set of temporal logic specifications representing mission goals and problem constraints. The key innovation includes new scalable algorithms for synthesizing distributed plans---receptive to the limitations imposed by perception---for multiple agents in partially observable environments withprovable guarantees, and new methods for real-time revising of agents plans to adapt to changing environments. Thrust III closes the loop by using information from planning and communication to guide and achieve task-aware, active perception. Its key innovations include new algorithms to improve perception with active coordination in large teams of robots with limited communication and active choice of sensing actions as well as new methods to further filter out irrelevant data for perception based on task-driven insights from planning. We will evaluate the resulting theory and algorithms with a range of experiments, from low-fidelity simulations to high-fidelity simulations to (limited) hardware experiments using an example scenario of disaster search and rescue.The expected outcomes of this work include theory, algorithms, and insights needed to develop new, scalable decentralized perception and planning approaches that are aligned with the emerging characteristics of Naval systems: highly autonomous with wide range of capabilities delivered in contested and uncertain environments with noisy and incomplete data sources.They have the potential to reduce the burden on human analysts, reduce the required communications bandwidth, improve the resilience and agility, and enable more rapid and effective decision-making for the next generation of highly autonomous systems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 19, 2018
- Source ID
- N000141812829
Entities
People
- Lu Feng
Organizations
- Office of Naval Research
- United States Navy
- University of Virginia