Variable Fidelity Simulation and Replay for Unmanned Autonomous Ground Vehicles 2019-01-1074

Robotics and automation researchers rely heavily on simulation and modeling as testing and development tools, to offset the time and effort required working with complex physical systems, especially those requiring large vehicles and expansive environments, such as unmanned military vehicle research. With increasing system complexity, unmanned vehicle simulators are becoming almost equally complex, potentially growing cumbersome and almost as time and resource intensive to use as testing in a physical environment. This ever-increasing system complexity necessitates a different, variable approach to simulation: using both low- and high-fidelity simulation environments to speed up algorithm development of high-level vehicle behaviors in a light-weight, easy-to-setup simulator, without sacrificing fidelity where it’s most needed: low-level motion control, mobility and perception. In this paper, we describe a variable fidelity simulation methodology we have employed to develop, integrate and test low-level motion and perception as well as higher-level tactical vehicle behaviors for autonomous ground vehicles in a government-sponsored project, resulting in successful field demonstrations. To enable autonomous vehicles to act as part of an infantry squad, performing similar tasks to those done junior squad members, we used two different simulation environments: 1) a ROS-based, in-house high-fidelity simulator for low-level mobility and perception, including local path planning, trail and person detection; and 2) SimJr., a Java-based, low-fidelity, lightweight simulator for development and testing of high-level tactical vehicle behaviors, such as maintaining a formation within a fire team, or performing a rear security task. Additionally, we examine advantages of system-wide ROS-based logging and replay to capture inputs at the module boundary so that performance on both sides can be tested independently, and usage of a hardware-in-the-loop technique to test a tablet-based user interface with simulated vehicles and people which allows testing UI functionality on a physical device without the effort and expense of using a vehicle.