Title: Coordinating Heterogeneous Teams for Urban Search and Rescue

The mission of Urban Search and Rescue (USAR) has drawn significant interest in robotics. Autonomous entities must be able to share knowledge efficiently to address visibility and collaboration challenges in a complex environment shortly after structural collapse catastrophes. In this thesis, we present methods to coordinate a rescue team consisting of members with different domain knowledge and capabilities. We first formalize a framework for modeling agents-environment interaction that allows for arbitrary origins of heterogeneity, from which we identify two configuration instances generalizable to common real-world tasks. We use them to develop a series of USAR environment simulators with concerns about the trade-off between fidelity and sample efficiency. Under these preparations, we propose a multi-agent reinforcement learning algorithm to tackle USAR. We adopt graph attention, in a novel manner, to fuse information perceived across agents and exploit structural knowledge regarding the environment. We apply action-dominant agent indexing to benefit from the power of parameter sharing, while still allowing agents to have different behavioral traits. We show that our proposed approach outperforms previous state-of-the-art literature by 40% to 120%. In addition, we develop hierarchical planning-based agents that mimic human behavior. We conduct imitation learning over faux human trajectories and demonstrate the improvement over purely online reinforcement learning. Ultimately, we show that the feature distribution between synthetic and human data is sufficiently close, so that the former can be used to forecast human performance. We observe that the inference error can be reduced by half when transformer-based predictors are augmented with a synthetic dataset.

Katia Sycara (chair)

Changliu Liu

Tejus Gupta