Our research involves evolving AI into proactive agents capable of complex reasoning, planning, and task execution. We are building systems capable of interpreting high-level user instructions, decomposing them into actionable steps, and autonomously interacting with software systems to achieve specific goals. Our research delves into optimizing the decision-making process of these agents, ensuring they can handle multi-step workflows with precision while adapting to the users’ unique environments and preferences.

A key focus of this theme is personalization and context awareness. We strive to create agents that act as reliable digital assistants, capable of organizing information and facilitating seamless interactions between different services. By grounding AI responses in relevant, real-time data and maintaining long-term memory of user interactions, we ensure that our agents provide support that is not just accurate, but deeply attuned to the individual’s daily life and specific needs.

We envision a future where intelligence extends beyond the screen to interact meaningfully with the physical world. Our research in embodied AI focuses on equipping devices—from smartphones and home appliances to robotic systems—with the ability to perceive, understand, and navigate 3D environments. Our focus lies in developing technologies that enable machines to recognize objects, interpret spatial relationships, and manipulate their surroundings safely and effectively, bridging the gap between digital algorithms and physical reality.

The research explores ways to enable devices to learn from physical interactions, improving their adaptability to unstructured environments and dynamic scenarios. Whether it is a home robot assisting with chores or a smart appliance optimizing its operation based on environmental cues, our goal is to deliver tangible, physical assistance that enriches the human experience in the real world.

We are advancing the frontiers of intelligence by developing multimodal foundation models that seamlessly process and generate language, code, vision, and speech. Our research encompasses the entire lifecycle of model development, from data curation and architectural design to efficient on-device deployment. By optimizing these models for various computing environments, we aim to provide powerful, real-time AI experiences that are both accessible and privacy-preserving.

This technology powers a wide array of applications, enabling natural, human-like interactions where users can converse with their devices, edit media through simple commands, and receive context-aware assistance. Our goal is to create a versatile intelligence that not only enhances individual productivity but also transforms how users create and consume content across different platforms.