Learning Mental Models of Experts in a Simulated Search and Rescue Scenario

Abstract

Search and Rescue is a task where the rescuers need to be cognitively agile, strategically consistent and efficient to save as many trapped victims as possible. However, in such a dynamic scenario, the rescuers’ mental models may be outdated, and it may be difficult to coordinate with other rescuers under a time constraint and cognitive overload. In this thesis, we propose to develop agents based on Machine Theory of Mind (MToM) to infer the beliefs, intentions, and desires of the rescuers from their observations and actions. By generating a mental model, an agent can intervene when it detects a rescuer might act based on a false belief. We also study approaches on using data from a different map to learn robust neural agents. Finally, we study the paradigm of imitation learning to learn policies from expert trajectories, and propose a sample efficient method over existing baselines.