Deep learning and variable autonomy for mobile robot navigation

Abstract

This paper addresses the problem of safe and efficient navigation of mobile robots in extreme environments such as Search and Rescue (SAR) and Urban Search and Rescue (USAR) missions or other forms of remote robotic tasks. A shared control module is utilized and expanded upon by introducing a Deep Neural Network model (DNN), as well as a Machine Leaning (ML) model. The models were trained to predict what is known as probability of suffering, a concept borrowed from robot vitals, which is a state-of-the-art performance degradation metric that this paper is heavily inspired from. The models use sensor, or close to the sensor level, data as inputs and make inferences by circumventing the need for mathematical calculations of robot vitals. These inferences are then fed to an arbitration function 𝑎(⋅) that dictates control allocation between human and robot. The models presented did not demonstrate a significant improvement compared to the shared-control module that was presented in our previous work (i.e. static value arbitration). More specifically, experimental evaluation did not show a clear performance increase in terms of safety and task completion time. Other measurements however (e.g. the behavior of the arbitration function), show that this work can act as a proof of concept for robot vitals and AI integration in shared control systems, offering possibly a versatile and scalable solution for arbitration policies.