Human-Agent Collaboration under Incomplete Information: Intent, Communication, and Planning

Shenghui (Vivian) Chen
University of Texas at Austin, Computer Science

Time: 2025-04-09, 12:00 - 13:00 ET
Location: Olsson 105 and Zoom

Abstract To collaborate effectively with humans, autonomous agents—whether embodied (e.g., robots) or virtual—must communicate and plan strategically under incomplete information. While most work, if not all, focuses on command execution under human authority hierarchy (e.g., recent development of LLM agents), I envision agents as equal partners rather than rule-following subordinates. In this talk, I present algorithms for human-agent coordination that integrate communication, planning, and reasoning over partner intent. Specifically, I will discuss: (1) single-step intent expressed through natural language, (2) multi-step intent as reward augmentation, and (3) unified action-communication planning with learned human perception dynamics. Through human subject experiments in two testbeds—Gnomes at Night and CoNav-Maze—we evaluate these methods on task metrics like coordination efficiency and user experience metrics like cognitive load.

Bio Shenghui Chen is a fourth-year Ph.D. student in Computer Science at UT Austin, advised by Ufuk Topcu. She earned her B.S. in Computer Science from UVA. She is interested in multi-agent interactions and human-agent collaboration, with a recent focus on developing planning and learning algorithms to enhance synergy between humans and autonomous systems.