Miss. Shenghui Chen

PhD

University of Texas at Austin

As intelligent agents become embedded in daily life—whether virtual, embodied, or wearable—the question shifts from acting alone to collaborating with humans. My research envisions agents not as tools, but as partners that understand intent, plan adaptively, and communicate effectively under incomplete information. Approach. I develop methods for intent inference, planning, and communication by combining algorithmic design with human-subject experiments. I introduced shared-control games [1] as a formalism for modeling asymmetries in control, perception, and knowledge between agents and humans. Inferring Human Intent for Joint Planning: Effective collaboration requires reasoning about human intent—commitments to future actions—based on diverse signals such as language [1], trajectories [3], or movement [5]. I use Bayesian belief updates and a Monte Carlo Tree Search-based planner to synthesize policies that account for inferred intent and uncertain dynamics. Efficient Agent-to-Human Communication: As agents increasingly outpace humans in information processing capacity, effective collaboration requires communication that is both informative and cognitively efficient. In a robot navigation setting [4], I modeled how humans update their understanding in response to robot visual communication. Embedded in a planner, this model enables robots to choose what and when to communicate. In a VLM setting [2], I developed annotation-free metrics based on information bottlenecks to evaluate and select grounded, task-relevant summaries—improving human accuracy by 61% and reducing response time by 76%. Looking Ahead. I aim to advance adaptive, human-centered agents through two directions: (1) studying how symbolic and sensory feedback across modalities influences human response, and (2) inferring intent from natural, unconstrained behavior using probabilistic models. These efforts extend to assistive robotics, AR/XR, and human-AI decision support.