T3DDART – Tractable Tiered Task Decompositions for Decentralized Autonomous Resource Tasking

The current state-of-the-art for autonomous unmanned vehicles is limited in comms-denied situations, where operators cannot assist with coordination and situational reassessment.  Significant algorithmic development is needed for comms-challenged autonomous operations to achieve parity with operator-driven or manned equivalents, as autonomous agents are not currently equipped to anticipate and resolve complex collaborative decisions adaptively and intelligently.  In a comms-challenged dynamic environment with diverging situational awareness and unreliable/unavailable inter-agent communication links, providing a generalized framework for autonomous agents to successfully execute missions flexibly and reactively is a profound challenge. In this proposal, we describe an R&D plan to create a framework for decentralized multi-agent task execution capable of operating in a variety of communication regimes, including 1. full EMCON (emissions control) or otherwise comms-denied environments, and 2. intermittent environments with sufficient bandwidth to perform limited deconfliction of situational understanding and/or tasking.  The baseline approach we lay out here is focused primarily on pre-mission policy generation that will enable reactive execution of multi-asset missions in fully comms-denied environments.  However, we additionally contend that a generalized system for decentralized coordination should not only be flexible enough to operate under denied comms, but also to leverage any level of available communication, at any spatial or temporal resolution, to ensure optimal collaborative mission success. The team- and Deep Neural Net (DNN)-augmented MacDec-POSMDP (Macro-actions Decentralized Partially Observable Semi-Markov Decision Process) approach serves as a top-level policy generator that adds uncertainty and information awareness to help decide how/when to leverage existing autonomous behaviors that may be implemented using state machines, reactive control, behavior trees, etc.  Under this effort, Scientific Systems Company, Inc. (SSCI) partnered with the MIT Aerospace Controls Laboratory (ACL) under Prof. Jonathan How, will adapt the listed components to the proposed T3DDART DRM (design reference mission), and will study the benefits and drawbacks of the proposed approach.  In Phase I, we will aim to demonstrate the feasibility of our pre-mission policy-generating approach in a simulated comms-denied (or limited-comms, depending on customer interest) scenario.  Phase II will focus on demonstration of our decentralized autonomy approach working across multiple scenarios with realistic agent platforms and system model constraints (such as sensor range, communication throughput and energy considerations).  If there is customer interest, we will additionally develop onboard models of spatial and temporal communications availability, to enable the system to flexibly adjust its execution to leverage additional bandwidth for improved performance.