Event
Ph.D. Research Proposal Exam: Ziyao Wang
Friday, October 30, 2026
10:00 a.m.
AVW 2328
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Ziyao Wang
Committee:
Professor Ang Li (Chair)
Professor Sennur Ulukus
Professor Heng Huang
Date/time: Friday, October 30, 2026, at 10:00 AM
Location: AVW 2328
Title: Collaborative Foundation Model System for Embodied AI on the Edge
Abstract: Foundation models offer a promising path toward embodied intelligence, extending AI from virtual environments into the physical world. However, operating these models on real-world devices introduces challenges in on-device training and deployment, data privacy, and edge--cloud collaboration. Addressing these challenges requires systems that connect the capabilities of foundation models with the resource constraints and evolving demands of embodied interaction.
My research investigates collaborative training, collaborative inference, and foundation-model agents, with the goal of enabling efficient and adaptable embodied intelligence. My completed work spans federated learning for large models, collaborative decoding, edge--cloud model collaboration, and data engines for embodied AI. These efforts establish a foundation for learning from distributed data, coordinating models with different capabilities, and supporting embodied learning through data collection and processing.
Building on this foundation, I propose to develop a collaborative machine learning system that enables generalizable, self-evolving embodied intelligence while minimizing computation on edge devices. The proposed research will pursue three connected directions. First, I will investigate embodied training data, including how interaction data can be collected, curated, and used to support generalization across tasks and environments. Second, I will develop foundation-model agents that integrate perception, reasoning, and action, and improve through accumulated experience and feedback. Third, I will explore agent-based edge-cloud collaboration that dynamically coordinates local execution and cloud assistance according to task requirements and resource availability. The system will be evaluated in terms of task performance, generalization to unseen tasks and environments, improvement through experience, and computational and communication costs. This research aims to establish system principles and mechanisms for bringing capable, adaptable embodied AI to resource-constrained devices.
