RAAMM (Robust Adaptive Autonomy for Multi-agent Maneuvers)

The overall objective of the RAAMM project is to advance technologies for multi-agent autonomy in complex and contested environments. The project is split into 4 tasks.

1. Creating a simulation environment for autonomous navigation in semi-structured environments.
2. Validating algorithms for autonomous navigation.
3. Developing methods for tactical multi-agent planning.
4. Solving the “Finding Waldo” problem, where a zoom camera device tracks an entity of interest in a partially occluded establishment; for example, tracking the movement of a person in a building through images that are captured via glimpses of the person through windows, or tracking a vehicle with aerial cameras through a wooded area.

ARI has been working on task 1, while the Army Research Lab and other University of Illinois faculty and graduate students tackle the other tasks.

• Girish Chowdhary’s group from in Agricultural and Biological Engineering
• Huy Tran from Aerospace Engineering
• Ramavarapu Sreenivas from Industrial & Enterprise Systems Engineering

We used Unreal Engine, a popular gaming engine, which provides an excellent platform upon which to develop full end-to-end simulations including vehicle physics, interactions with the environment, and sensor suites that mimic real systems. The Carla plug-in supports configurable vehicles with a variety of mounted sensors, whose data may be streamed and saved.

We started by building large landscapes and importing terrain height maps using TerreSculptor.

We improved the diversity and accuracy of the maps by synthesizing new material textures and importing a better mix of rocks, debris, and foliage from Quixel Megascans, a 3D asset library. We edited a model of the TerraSentia robot by EarthSense in Blender and then imported it into the simulation, where we adjusted vehicle dynamics by experimenting with various parameters such as mass properties, suspension, and friction.

We finally delivered 2 maps for the TS bot to navigate: a dense forest map with a dirt path and a mountainous grassland map with sparse trees and rocks. The forest map provides mud patches, potholes, and solid objects for the bot to traverse. The grasslands map has fewer obstacles but limits the bot visibility due to the tall grass which adds visual clutter to the camera sensors.