Situational Risk Assessment Design for Autonomous Mobile Robots.

The emerging autonomous mobile robots promise a new level of efficiency and flexibility. However, because these types of systems operate in the same space as humans, mobile robots must cope with dynamic changes and heterogeneously structured environments. To ensure safety, new approaches are needed that model risk at runtime. This risk depends on the situation and is therefore a situational risk. In this paper, we propose a new methodology to model situational risk based on multi-agent adversarial reinforcement learning. In this methodology, two competing groups of reinforcement learning agents, namely the protagonists and the adversaries, fight against each other in the simulation. The adversaries represent the disruptive and destabilizing factors, while the protagonists try to compensate for them. The situational risk is then derived from the outcome of the simulated struggle. At this point, the system's Digital Twin provides up-to-date and relevant models for simulation and synchronizes the simulation with the real asset. Our risk modeling differentiates the four steps of intelligent information processing: sense, analyze, process, and execute. To find the appropriate adversaries and actors for each of these steps, this methodology builds on Systems Theoretic Process Analysis (STPA). Using STPA, we identify critical signals that lead to losses when a disturbance under certain conditions or in certain situations occurs. At this point, the challenge of managing the complexity arises. We face this issue using training effort as a metric to evaluate it. Through statistical analysis of the identified signals, we derive a procedure for defining action spaces and rewards for the agents in question. We validate the methodology using the example of a Robotino 3 Premium from Festo, an autonomous mobile robot. [ABSTRACT FROM AUTHOR]