Navigation Improvements for Maneuvering Between Limited Environments (NIMBLE)

Self-driving convoys promise the ability to move large payloads with reduced human oversight. Human operators drive vehicles followed by a tail of autonomous systems. Unfortunately, limitations exist within the navigation stack of contemporary autonomous vehicles that inhibit their ability to safely navigate through narrow passages and around hazardous terrain obstacles. Failures in the capabilities of these systems require that human operators assume control of these vehicles and manually steer them through and around these challenging scenarios. Handing off control of these vehicles can incur significant delays and is antithetical to the motivation behind self-driving vehicles. Charles River Analytics proposes Navigation Improvements for Maneuvering Between Limited Environments (NIMBLE) an approach to identifying the key components that result in the failure of these navigation systems. First, we will use the Government-owned Robotic Technology Kernel and ProjectGL to simulate environments that cause navigation faults in uncrewed ground vehicle (UGV) platforms. These simulations serve to interrogate the capability of the individual components of the navigation stack. Then, we will use this simulator to evaluate the potential improvements that alternate sensor configurations may provide. Finally, we will train and integrate a deep learning¢based perception system to identify and segment hazardous terrain features.