GRAPPLE: Generalizable Robust & Adversary-adaptive COA Planning via game-Play & Learning Environments
Abstract
Project AbstractWe propose a new paradigm in AI-based mission planning and execution. Our research team comprises University at Buffalo (as lead) and Palo Alto Research Center or PARC (as sub-contract). Our overarching research objective is to develop AI architectures to: 1) generate and rank sets of competing COAs for missions in real-world adversarial settings, 2) use human game-playing environments for evaluating adversary behavior and updating COA in response, and 3) identify constrained changes to COA for misguiding adversary s assessment of the mission state. With growing complexity of military and humanitarian missions, AI-based or AI-aided decision-support can uniquely enhance mission-planning speed and capacity compared to pure human decision-making. However, the ability to support the planning of complex missions at scale including generating, ranking, and dynamically updating COA in the presence of smart adversaries remains out of reach of existing AI frameworks. To addressthis need, our proposed technical concept, #GRAPPLE: Generalizable Robust & Adversary-Adaptive COA Planning via Game-Play & Learning Environments# can be broken down into three major innovations or #Themes#: 1. To implement geometric and graph learning architectures that can interpret high dimensional and multi-modal environment information and learn robust AI models that embody an interpretable pool of COAs and their ranking in terms of mission-focused fitness.2. To identify and test methods for using human demonstration of real-time strategy (RTS) like games in order to design adversary models, and predict adversary behavior from movement, capability and activity observations.3. To identify and test reward shaping schemes and error-inducing multi-UxV cooperation rules that modify the design and primitive-based execution of COA to misguide adversaries. To provide context and an evaluation platform for applying GRAPPLE, we will use the following prototypical application: search and extract mission by a team of 50-100 unmanned ground and aerial vehicles (UGVs and UAVs) operating over a large semi-urban area that includes varying terrain and active adversarial units. GRAPPLE provides the mission-focused AI to help command suchan operation. The proposed GRAPPLE concept will be extensively evaluated in simulation, and targeted set of physical experiments. We expect our proposed generalizable, interpretable and dynamically adaptive AI-based planning paradigm will allow the U.S. Navy and DoD to preserve their strategic advantage in complex missions.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 21, 2023
- Source ID
- N000142412003
Entities
People
- Souma Chowdhury
Organizations
- Office of Naval Research
- Research Foundation for the State University of New York
- United States Navy