Towards personalized communication between humans and their assistant systems

Communication informs various facets of human-autonomy collaboration such as task efficiency, trust, and safety. Particularly so in use cases where humans interact with assistant systems (e.g. navigation assistants, guide robots, collaborative manufacturing). The current communication paradigms between humans and their assistant systems are direct and static, leading to robotic experiences. Learning-based techniques can be used to improve the aforementioned facets by exploiting dynamic, implicit communication cues and enhancing direct communication strategies. To this end, we aim to use learning-based techniques to improve human-autonomy communication. We enhance human modeling in collaborative tasks to inform unsupervised inference and evaluate our methods with human-in-the-loop validation. Specifically, we model human-autonomy interaction as Markov Decision Processes (MDP) and build assistant policies using Reinforcement Learning. The MDP formulation is then used to train models such as generative Variational Autoencoders and graph neural networks that capture and predict human traits and preferences among other human-behavior. These predictive models enable (semi-)autonomous assistant systems to adapt and cooperate with their human counterparts. We evaluate our proof-of-concept systems using user trials, with particular emphasis on in-car driving scenarios. Our experiments provide important insights into interactions between humans and their assistant systems. We show how implicit cues provided by humans during interactions can be capitalized on to improve learning-based assistants. User interactions with our proof- of-concept systems aid us in providing recommendations for the design of future assistant systems, particularly in the case of real-time speed advisors. We also provide recommendations for direct communication strategies and showcase their effects on user experiences. Overall, the work presented in this dissertation shows that learning-based techniques for human behavior modeling that lead to robust human-behavior prediction are beneficial in aiding user experiences while also leading to improved performance of the human-assistant team.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2024

Author dc:creator

• Hasan, Aamir

Contributors dc:contributor

• Driggs-Campbell, Katherine
• Driggs-Campbell, Katherine Rose
• Dong, Roy
• Karahalios, Karrie
• Varshney, Lav

• Applied Machine Learning
• Human Assistant Systems
• Human Robot Interactions
• Adaptive Communication
• Human-in-the-loop Evaluation

Statement dc:rights

• Copyright 2024 Aamir Hasan

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/127231

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F127231