Personal Homepage

Overview

Our research aims to advance general-purpose artificial intelligence by building systems that can perceive, reason, and act in open-world environments.

We focus on the integration of embodiment, multimodal foundation models, and cognitive-inspired learning, with the goal of movingbeyond static pattern recognition toward adaptive, interactive, and generalizable intelligence.

A central theme of our work is bridging the gap between data-driven learning and structured cognition, enabling AI systems to operate robustly under distribution shift, limited supervision, and long-horizon tasks.

Research Areas

Embodied Intelligence & Autonomous Agents

We study how agents interact with complex environments through perception, action, and feedback.
Key problems include:

• Long-horizon decision making

• Environment modeling and world representation

• Generalization across tasks and domains

Multimodal Foundation Models

We develop models that unify vision, language, and other modalities into a shared representation space.
Our focus includes:

• Cross-modal alignment and reasoning

• Multimodal generation and interaction

• Scaling laws and emergent capabilities

Cognitive-Inspired Multimodal Learning

We explore mechanisms inspired by human cognition to improve learning efficiency and robustness:

• Memory and retrieval-augmented models

• Structured reasoning and compositionality

• Multi-step inference and planning

Learning with Limited Supervision

We design methods that reduce reliance on labeled data:

• Self-supervised and weakly supervised learning

• Few-shot and in-context generalization

• Adaptation in open and dynamic environments

Selected Projects

Our work is supported by major national and international research programs, including:

• National Key R&D Program on Next-Generation AI

• National Natural Science Foundation (Key Projects)

• International collaborations with leading institutions (e.g., CMU)

These projects focus on generalist agents, multimodal reasoning, and scalable intelligent systems.

Publications & Impact

• 150+ publications in top-tier venues (TPAMI, IJCV, NeurIPS, ICML, CVPR, ICCV, etc.)

• 20,000+ citations on Google Scholar

• 21 ESI Highly Cited / Hot Papers

• Clarivate Highly Cited Researcher (2019–2024)

Research Philosophy

We prioritize:

• Problem-driven research over incremental improvements

• Generality and robustness over benchmark-specific gains

• Depth and originality over short-term output

Our goal is to develop principled and scalable approaches to intelligence, rather than isolated techniques.

Open Problems (We Are Hiring)

We are actively looking for students interested in tackling challenging problems such as:

• General-purpose embodied agents

• Multimodal reasoning beyond pattern matching

• World models and environment understanding

• Scalable learning under limited supervision

Collaboration

We welcome collaborations with academic and industry partners on:

• Embodied AI systems

• Multimodal foundation models

• Autonomous agents in real-world applications