Resilient Multi-Agent Perception and Planning in Dynamic Domains
Abstract
This project tackles the need to create high-confidence,high-reliability, minimal-supervision autonomous systems that canunderstandand act in high op-tempo missions. These systems should beable to deal with dynamically evolving environments that requirereal-time extraction and processing of task-relevant information fromheterogeneous tactical and intelligence streaming data obtained bysensors, machines, and humans; build hierarchical representations ofthe environment that allows reasoning at different levels ofgranularity; integrate perception and planning to supportcollaborative and distributed task execution with formal performanceguarantees; reliably transition from offline training domains toop-tempo deployments in the field; efficiently predict actions of redforces in the field, reason about significant changes overpre-computed plans, perform decision-making over short timespans, anddeal with environments unlike those seen in simulation; andorchestrate the team actions to enable robust and distributedintelligence under partial and spatially fragmented information,different processing, sensing, communication, and controlcapabilities, and asynchronous decision-making.The proposed technical approach addresses head-on key objectivespertaining to hierarchical, multi-scale, and resilient situationalawareness in dynamic environments; the integration of perception andplanning; distributed methods for perception and prediction; anddistributed planning and control algorithms for multi-agent teamsdeployed in dynamic environments. The synergistic research program isorganized along four thrusts. Thrust 1 focuses on hierarchical andresilient reasoning about dynamic environments at multiple temporaland spatial scales informed by metric and semantic information. Thrust2 develops integrated hierarchical perception and planning to supportdecision-making at multiple levels of spatial, semantic, and temporalabstraction, providing the foundation to tacklecommunication-dependent multi-agent scenarios. Thrust 3 buildssituational awareness and hierarchical abstractions from streamingdata for multi-agent teams to enable operation in a distributedfashion. Thrust 4 complements the distributed perception techniques inThrust 3 by focusing on the design of scalable distributed algorithmsfor multi-agent planning and optimal control under communication,dynamics, and perceptual constraints.The proposed research will significantly enhance the Navy scapabilities in intelligence, surveillance, and reconnaissance (ISR)missions by advancing tractable computational methods for flexible andresilient learning, sharing, and reasoning. Our approach will enableenvironment representations for multi-agent situational awareness aswell as autonomous execution of complex multi-agent missions. RAPIDwill endow agents with the capabilities to build compactmetric-semantic models that integrate localized and dynamic perceptioninformation into a representation that can be shared with otheragents. This will enable robotic teams to attain a comprehensive,global, and accurate understanding of the environment, and to plancollaboratively with other agents within a unified theoreticalframework that does not require ad hoc decomposition or restrictiveassumptions on the separation of perception and planning. The projectwill establish fundamental theory and algorithms for reasoning aboutdynamic scenes, managing the impact of uncertainty in decision-makingand planning, leveraging multi-source, multi-scale perception andhierarchical planning for fast anddistributed decision-making atmission op-tempo, and resilient operation and deployment outside theoriginal training domain.This abstract is approved for public release.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 12, 2023
- Source ID
- N000142312353
Entities
People
- Jorge Cortés
Organizations
- Office of Naval Research
- United States Navy
- University of California, San Diego