Research

Research Vision

We aim to advance physical AI that can perceive, reason about, and act in the real world. Our research connects multimodal perception, spatial intelligence, and embodied learning to turn observations into meaningful action. By learning from visual experience, simulation, and real-world interaction, we aim to build intelligent systems that adapt to unfamiliar environments and work alongside people.

Research Directions

Perceive the World

Multimodal Perception

Physical intelligence begins with sensing. We study how machines can learn meaningful representations from visual and multimodal experience, recognize what matters, and connect observations to language. Our goal is perception that remains useful across changing viewpoints, environments, and tasks.

• Open-world representation learning
• Multimodal grounding and recognition
• Robust perception across environments

Understand the Scene

Spatial Intelligence & World Models

Useful perception must become an understanding of the surrounding world. We investigate representations that connect geometry, semantics, and dynamics, allowing an agent to locate itself, remember its surroundings, and reason about possible changes. This direction connects 3D scene learning with spatial memory and predictive world models.

• 3D and 4D scene representations
• Spatial memory and reasoning
• Predictive and interactive world models

Move into Reality

Embodied Intelligence & Action

We aim to close the loop between seeing, understanding, and doing. Our research explores agents that connect human intent to action and learn from videos, simulation, and physical interaction. We work toward adaptable robot intelligence that can navigate, interact, and improve through experience in real environments.

• Vision-language-action learning
• Embodied navigation and interaction
• Learning across video, simulation, and reality

Research Videos