The goal of this project was to develop a generalizable framework for rigorous and systematic design of autonomy supervised by a team of interacting human operators, which enables the leveraging of human operators' adaptivity in complex scenarios while mitigating performance deterioration due to loss of situational awareness. Underwater search and rescue was used as a motivating application and case study.   

The project contributed to the development of novel robotic search and deployment algorithms that leverage the differences in the sensing at different altitudes in underwater search. The performance of the algorithm was rigorously characterized in terms of expected search time for a desired accuracy, and the algorithm was validated using an underwater robotics simulator in Gazebo. The algorithms were extended to multiple vehicles using environment partitioning. To this end, a novel algorithm that handles the tradeoff in learning the environment and equitably partitioning it were developed. The performance of the algorithm was characterized in terms of the cumulative regret, i.e., the gap between a partitioning metric achieved by the algorithm and the best possible partitioning metric. The search framework was extended to handle non-stationary environments through the development of novel algorithms for non-stationary multi-armed bandit problems. In addition to non-stationarity, these algorithms can also handle high uncertainty that leads to heavy-tailed sensing data.

A semi-Markov decision process (SMDP) framework was developed to assist human operators. Using a queueing theory setup in which the human operator is a server with their cognitive state-dependent performance, the proposed framework suggests to the human operator the fidelity level at which they should process the task, take rest, or skip the task. The structural properties of the optimal policy for such an SMDP were rigorously established. The framework was extended to learn human task processing model in real-time from limited human data using a Bayesian setting. The proposed algorithms were rigorously analyzed and numerically validated. A Gazebo-based human-supervised underwater search experiment was developed in which human cognitive state was estimated using eye-tracking data. Preliminary experiments were conducted to validate the assistive framework. However, due to COVID-19, human subject experiments were suspended and these experiments have been postponed.

The project also contributed to significant advances in several robotic platforms (e.g., gliding robotic fish and unmanned surface vehicles) with potential for application to underwater search and rescue. An LED-based underwater communication system for underwater robots, including hardware and algorithms, was developed and tested in underwater experiments. Furthermore, LED-based localization schemes for mobile robots were developed and evaluated in experiments involving a rover robot, which could be further extended to a 3D space including the underwater setting.

Collectively, the project resulted in 14 journal publications (10 published,  4 under review), 19 juried conference presentations and papers, 4 PhD dissertations, and 1 Master’s thesis. It provided training opportunities to 5 PhD students, 1 Master’s student, 7 undergrad students, and 1 K-12 teacher. Finally, the project contributed to various activities for outreaching to K-12 students and the general public, including engagement through MSU high school engineering institute, summer research opportunity programs,  MSU science festival, lab tours, and radio interviews.  

Last Modified: 12/28/2021
Modified by: Vaibhav Srivastava