Stochastic Distributed Optimal Dual Control: A Unified Framework for Decentralized Multi-agent Perception and Planning (SDODC)
Abstract
Sensing and reasoning about processes and objects in a large physical environment requires both collecting and processing large-scale spatio-temporal data, typically through teams of autonomous agents. The proposed research establishes a unified framework based on stochastic distributed optimal dual control (SDODC) for decentralized task aware perception and asynchronous planning for autonomous agents deployed to collect and process large amounts of complex spatiotemporal streaming data. Proposed AI perception algorithms are able to extract the most significant information from large amount of data streams. The SDODC framework will consist of a systematic approach for incorporating AI perception algorithms that will be demonstrated on multi-agent systems engaged in semantic planning, automatic target recognition, mapping of spatial-temporal fields, and obstacle avoidance. The project will overcome the technical challenges for decentralized perception and planning in dynamic environment by offering scalability to large number ofagents and compatibility to AI perception algorithms. The SDODC theoretical framework will be based on the novel theory on finite set statistics (FISST). The FISST enabled SDODC offers a unified formalism for perception and planning bytreating the AI perception output and the network of agents as multi-object spatio-temporal densities connected by a spatial map. Perception and planning densities are jointly computed for optimality.The PIs will derive distributed principles of optimality and certainty-equivalence properties. Mappings between mainstream AI perception algorithms on data streams and density will be established for functions to enable task aware perception. When implemented distributedly, decentralized dual-control solutions and decentralized kerneldensity estimation (KDE) principles will be derived to accommodate changing link structures and link quality among agents. The proposed work find the minimum amount of map information to be shared among communicating agents for a satisficing sub-optimal control solution. The relationship between the quality of the map being shared and the sub-optimality of the control solution will be rigorously established to provide guidance for semantic maps constructed by each agent to be shared with their neighbors. The PIs will validate SDODC theory and algorithms via physical and virtual experiments involving teams of autonomous underwater and aerial vehicles.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 25, 2019
- Source ID
- N000141912266
Entities
People
- Fumin Zhang
Organizations
- Georgia Tech Research Corporation
- Office of Naval Research
- United States Navy