OBJECTIVE: Develop algorithmic approaches to enabling robust control of autonomous unmanned ground vehicles operating in complex, unstructured environments, over low-bandwidth, high latency communication links. DESCRIPTION: This topic addresses the problem of robustly commanding and controlling unmanned ground vehicles operating in complex, unstructured environments. Current approaches to this task rely on dense scene reconstruction from a variety of sensor data such as LIDAR and video imagery. Scene representations are then relayed to a remote human operator who provides commands at varying levels of supervisory control (intelligent teleoperation). Challenges arise in scenarios where the available communication link allows only low bandwidth data transmission and/or exhibits high latency. In such scenarios, bandwidth limitations prevent rich scene representations from being transmitted from the vehicle to the operator in a timely manner. In addition, high latency may have a destabilizing effect, causing commands issued by the operator to lead to unsafe actions by the vehicle. This effect is exacerbated as the frequency of command inputs increases. Novel frameworks, and associated algorithms, are required to enable robust operations of autonomous unmanned ground vehicles operating in complex, unstructured environments, over a low-bandwidth, high latency communication link. Such approaches would provide an operator with sufficient information to make timely command and control decisions, even in harsh communication scenarios. They would also provide contingency-based assurance of system safety in the absence of timely command and control decisions. In addition, other functional relationships such as sensor costs need to be addressed since these costs are generally proportional to the level of autonomy or intelligence. Approaches to this problem may emphasize perception, vehicle control, or some combination of the two. In the perception domain, approaches to intelligent data compression and minimal scene representation are desired . Such approaches may condense raw sensor data into compact, human-recognizable primitives that can be efficiently transmitted over low-bandwidth communication links. These methods may be optimized for particular contexts (e.g. urban operations) to enable improved data compression, and they may also dynamically vary scene representation richness or complexity depending on available bandwidth. In the control domain, contingency-based control algorithms that ensure vehicle safety in the absence of operator inputs, or when provided with unsafe command inputs (perhaps due to the effects of latency) are desired. Again, such approaches may be optimized for particular contexts to enable improved performance. Methods that act as vehicle"co-pilots", which both ensure vehicle safety and attempt to predict operator intent, would be particularly useful . The output of this work is software that would be integrated with an existing autonomous vehicle(s) to yield measureable improvements in safety and operational speed compared to a baseline system, for the low-bandwidth, high-latency scenarios of interest. If successful, this work will have broad applications for autonomous and semi-autonomous military vehicle operations. PHASE I: The goal of Phase I is to investigate the feasibility of developing algorithms that can robustly handle huge variations in both bandwidth and latency while still maintaining levels of overall mission control that are comprehensible to a human operator. On the unmanned ground vehicle side, these variations could in the worst case scenario transition within milliseconds from sufficiently high bandwidth and low enough latency to permit as-needed transitions into tele-operator mode, down to no communications at all. Upon such transitions, the unmanned ground vehicle should demonstrate an ability to replace direct human command with an emulation of human command that consists of some hierarchy (possibly very simple) of specific mission goals and navigation rules. On the human side, the mission supervisor software should similarly use its knowledge of the intent of the unmanned ground vehicle to provide the best possible visual representation of its likely status, including for example explicit graphical methods for representing growing uncertainty. Feasibility of the proposed framework/algorithmic approach(es) may be demonstrated through modeling and simulation for initial developmental verification. The Phase I deliverables shall include a final report detailing the algorithms, theory, and initial performance data. In addition, the final report should testing options for very diverse exploration of changing bandwidth and latency during the operation. The addition of noise is desirable but not required. PHASE II: Phase II shall produce prototype software with focus on implementation and testing of the framework/algorithm(s) developed in Phase I. Phase II prototype verification shall be demonstrated through modeling and simulation with particular emphasis on proof-of-concept and performance. General testing shall execute options developed in Phase I concerning exploration of changing bandwidth and latency during operation. The prototype software system shall include: Fully documented framework/algorithms in open source with compatibility to ROS. Prototype demonstrations using Modeling and Simulation in a credible simulation environment, or an unmanned ground vehicle platform.. Final Report detailing all development, design, and testing to include performance metrics. PHASE III: Work in Phase III will focus on the transition of Phase II prototype software to a DoD relevant environment and platform(s). Specific DoD interface standards (e.g. RS-JPO Interoperability, JAUS, etc.) shall be implemented as necessary to permit adaption to various platforms. Potential commercial applications of this technology include commercial automotive. DUAL USE APPLICATIONS: Unmanned ground vehicles have applications in numerous civilian domains, including mining, autonomous driving, hazardous site inspection, site security, and others. It is expected that the software developed here would have direct application in a variety of such domains.