ACCLIMATE: Adaptive Coordinated Control of Intelligent Multi-Agent Teams
University of California, Berkeley

ACCLIMATE: Adaptive Coordinated Control of Intelligent Multi-Agent Teams

A Multi-University Research Initiative Funded by The Army Research Office

Partners

Mission Statement

We propose the design and evaluation of the adaptive hierarchical control of mixed autonomous and human operated semi-autonomous teams that deliver high levels of mission reliability despite uncertainty arising from rapidly evolving environments and malicious interference from an intelligent adversary. The design of architectures combining both hierarchical and heterarchical elements, the analytical foundations of interacting hybrid systems, the design of controllers for such systems that are robust against uncertainty, the management of rich sensory information from networked sensors among distributed and mobile teams; and the incorporation of human intervention in a mixed-initiative system are all key areas of our work. Additionally, the novelty of our approach is to explicitly take into account the need to adaptively replan missions to take into account environmental uncertainties and the deliberate malicious actions of a determined adversary. Our approach builds on the following research thrusts:

Thrust I: Architecture Design and Analysis for Dynamic, Adaptive Planning.

The architectures that we design will organically incorporate human intervention at all levels of planning and execution. Architectural design begins with an overall hierarchy featuring flexible team formation, task specification, pre-mission evaluation, and changes in goals, team composition, and communications during mission execution. In the rapid adaptive, dynamic replanning, it is an absolute necessity to have modules which can be composed interoperably on the fly when warranted by the actions of the adversary. A main drawback of traditional approaches of hybrid systems has been their extreme conservativeness of compositionality when designing intrinsically complex architectures.

We will address these issues through work on:

1. Abstractions for perception and action, and
2. Assume-guarantee reasoning and interface theory for compositionality.

Thrust II: Integration of Rich Multi-sensor Information into Virtual Environments for Incorporating Human Intervention in Mission Planning and Execution.

A key difficulty in the use of unmanned aerial vehicles (UAVs), unmanned ground vehicles (UGVs) and unmanned underwater vehicles (UUVs) is the difficulty in acquiring the rich sensory data gathered by the vehicles contributing to sensory-overload where teams of up to four warfighters are required to control a single robotic asset (either a UAV or a flight control system). We propose to handle the fusion of rich multi-sensor information over an unreliable network by developing new classes of algorithms combining recent work in omni directional vision, the extraction of graphical models from video sequences, and the joint rendering of simulated (synthetic) environments with multi-sensor (real) data. The research directions are: Adaptive hierarchical networks for acquiring and providing information, Extraction of 3D models from distributed video and other sensors networks, and Environments for human intervention and decision making.

Thrust III: Handling Uncertainty and Adversarial Intent in Adaptive Planning.

Two types of uncertainty pervade mission planning and execution: Probabilistic uncertainty having to do with environmental unknowns, such as weather, terrain data uncertainty, the probabilistic nature of failures of hardware or software, information attack, Adversarial uncertainty having to do with systematic attempts by an intelligent adversary (red-force) to defeat the mission. A key mathematical framework for the modeling of adversarial actions comes from the theory of games, and partially observable Markov decision processes and games. We will develop methods for; learning of adversarial strategies. We will develop teaming and game strategies to allow for defeating a dynamic adversary, that is one who changes his strategy, cost function, information atterns during the course of an engagement. Finally the strategy for the integration of the research of the three University teams is through a set of two or more scenario-based challenge problems involving intelligent adversaries on the extensive testbeds at the three partner institutions. The scenarios are responsive to battlefield scenarios as well as other national security needs such as hostage rescue, tracking of unfriendly forces, and homeland security needs.

Site managed by David Shim, hcshim@eecs.berkeley.edu