Research Statement

Building Embedded Physical AI Systems

• At the KAIST Embedded AI Lab, we build embedded physical AI systems that can sense, remember, learn, reason, plan, adapt, and act across cyber and physical environments under real-world constraints on computation, memory, energy, and connectivity.

• Our work is rooted in Embedded AI, which brings intelligent computation to embedded devices through on-device processing and collaboration with other devices, edge infrastructure, and the cloud. This provides a broad foundation for research on AI models, learning algorithms, and system software under resource constraints. Our work on Embedded Deep Intelligence contributes to this foundation by enabling deep learning models to run, adapt, and learn directly on resource-constrained platforms.

• Building on this foundation, our focus is closed-loop intelligence in physical environments: systems that use observations and experience to choose actions, evaluate their consequences, and adapt over time. This requires coordinating AI models, sensing pipelines, memory, runtime execution, and physical interaction so that intelligent behavior remains useful and dependable as conditions change.

Why Embedded Physical AI? 

• AI can become more useful in everyday life through devices that interact directly with people and the physical world: robots that assist us, wearables that understand our needs, vehicles that respond to their surroundings, and machines that adapt to changing conditions. These applications require intelligence that can operate locally, respond on time, learn from experience, and remain dependable throughout deployment.

• Achieving these capabilities raises fundamental research questions. Embedded devices operate with limited computation, memory, energy, and connectivity, while physical interaction introduces deadlines, uncertainty, and changing environments. Decisions about what to sense, what to remember, when to learn, and how much computation to perform directly affect what an agent can understand and do. Learning algorithms, computational execution, and physical behavior must therefore be studied together.

• Embedded Physical AI provides a long-term research direction at the intersection of artificial intelligence, computer systems, and physical interaction. Its open problems include retaining useful knowledge within bounded memory, adapting models without losing existing capabilities, allocating computation under changing resource budgets, and coordinating timely actions across devices. Progress can enable more responsive and personalized assistance, reduce unnecessary communication and exposure of sensitive data, and support reliable operation when connectivity is limited.

Why KAIST Embedded AI Lab? 

At the KAIST Embedded AI Lab, you can explore these questions through research spanning both AI models and systems. We pursue computation-embedded intelligence: understanding how an agent’s computational state and execution structure shape its ability to perceive, remember, learn, reason, and act.

We are looking for researchers who want to help develop this direction and bring their own questions to it.

• Make an important problem your own. Explore how neural memory should retain and retrieve experience, how agents should learn within limited resources, or how execution should adapt to the demands of a physical task. We encourage researchers to identify the assumptions behind existing approaches and develop questions that can grow into a coherent research direction.

• Build depth and connect across the computing stack. Your starting point might be foundation models, learning algorithms, operating systems, runtimes, compilers, or embedded hardware. Our research connects these areas, creating opportunities to develop a strong specialty and understand how your ideas enable advances elsewhere in the system.

• Turn your ideas into systems you can experiment with. Develop a model, implement a memory mechanism, build a runtime, or integrate an intelligent agent on a device. Study how your design changes learning capability, responsiveness, energy use, and dependable behavior. Working systems make discoveries tangible and reveal new questions that emerge during deployment.

• Grow into an independent researcher through collaboration. We value open discussion, rigorous experiments, shared learning, and mutual support. Our goal is to help one another develop the judgment to choose meaningful problems, the technical depth to solve them, and the clarity to communicate the results.

Our long-term vision connects individual intelligent devices with networks of agents that learn and act together. Your research can contribute a building block to that vision while developing ideas and skills that carry into future problems.

If you want to understand how intelligence works under real-world constraints—and enjoy creating the models and systems that make it possible—we invite you to explore these questions with us.

Our Research Approach and Areas 

• Our research spans both AI models and systems. Through full-stack model–system co-design and co-optimization, from foundation models and learning algorithms to operating systems, runtimes, compilers, and hardware, we make advanced AI deployable, adaptive, efficient, and dependable on real-world devices.

• We treat an agent's computation — its execution structure, memory state, and adaptation process — as part of its embodiment, not a layer separate from the physical world it acts in. Under this view, making AI work in the physical world is not primarily a modeling problem or a systems problem in isolation, but the problem of co-adapting computation and physical behavior under shared, real-world constraints on compute, memory, energy, and connectivity. We call this computation-embedded intelligence, and it is the single thesis that organizes everything we build: we study how choices about execution, memory, sensing, and learning shape what an agent can perceive, remember, decide, and do as its resources and environment change. 

Our central research question is:

• • How can we design physical AI agents whose computation — model execution, memory, and adaptation — is treated as part of their embodiment, so that models, memory, and actions jointly adapt to learn and operate reliably on embedded devices under explicit constraints on memory, latency, and energy as resources and environments change? 

Our research spans three areas:

• Execution Layer (AI Systems, Operating Systems, Runtimes, and Compilers): System software for efficient, portable, and dependable AI execution across heterogeneous embedded hardware.

• Model Layer (Device-Native Foundation Models and Agents): Multimodal foundation models and autonomous agents co-designed from the ground up with real-world device constraints.

• Memory Layer (Neural Memory, On-Device Learning, and Continual Adaptation): Neural memory and learning systems for knowledge retention and retrieval, personalization, and continual adaptation directly on resource-constrained devices.

Across these areas, we develop reusable algorithms and system mechanisms, and study how their integration enables dependable physical interaction.

Toward Distributed Physical Intelligence:

• • Our long-term vision extends embedded physical intelligence from individual devices to collaborating agents. Building on our research in AI execution, device-native models, and neural memory, we aim to enable agents to share knowledge, coordinate computation and actions, and learn together under resource, communication, and privacy constraints. This extends computation-embedded intelligence across devices, connecting dependable local behavior with collective capabilities in the physical world. 

Our Goals and Values  

• Research Excellence. We advance electrical engineering, computer science, and artificial intelligence through original ideas, rigorous investigation, and useful software and hardware. We share our findings and engage with the research community.

• Meaningful Impact. We address real problems and develop technologies that improve people’s lives. We communicate our research clearly so that others can understand, evaluate, and use it.

• Collaboration and Learning. We support one another, welcome different perspectives, and work across disciplinary boundaries. We view challenging research as a shared effort that develops both knowledge and researchers.