Research Statement

What are Embedded AI and Embedded Physical AI? 

• Confluence of Embedded Systems, Artificial Intelligence, and the Physical World 

• Along with the rapid development of computing technologies and the widespread deployment of embedded, mobile, wearable, robotic, vehicular, and IoT devices, an enormous amount of data is now generated directly from the physical world. These devices continuously observe people, objects, environments, machines, and biological signals through sensors such as cameras, microphones, IMUs, radars, LiDARs, tactile sensors, and physiological sensors. As embedded and IoT devices become increasingly distributed and pervasive, they are no longer merely data collection endpoints; they are becoming active computing platforms that can sense, learn, reason, and act in real time. 

• Meanwhile, Artificial Intelligence (AI), broadly defined as intelligence exhibited by machines, has advanced rapidly through breakthroughs in machine learning and deep neural networks (DNNs). AI has demonstrated remarkable success in solving complex problems in autonomous driving, natural language processing, healthcare, robotics, smart environments, industrial automation, and scientific discovery. However, many of these intelligent capabilities are still primarily designed for cloud servers or high-performance computing platforms, while the data and the actions often originate from physical devices operating under strict resource constraints. 

• This mismatch creates a strong demand to integrate embedded systems, artificial intelligence, and physical-world interaction. This gives rise to Embedded Artificial Intelligence (Embedded AI): a computing paradigm in which intelligent tasks are performed close to or directly on embedded devices, rather than relying solely on remote cloud servers. Embedded AI enables devices to process data locally, reduce communication cost, improve privacy, respond with low latency, and adapt to changing environments.

• Building on this paradigm, we define Embedded Physical AI as the next stage of Embedded AI: AI systems that can sense, learn, reason, and act in the physical world under strict memory, latency, energy, reliability, and hardware constraints. Embedded Physical AI is not simply about deploying smaller AI models on devices. Rather, it requires rethinking AI algorithms, system architectures, runtime mechanisms, sensing pipelines, learning methods, and deployment strategies so that intelligence can operate continuously and reliably in real-world embedded systems. 

• Embedded Artificial Intelligence and Embedded Deep Intelligence 

  • Embedded Artificial Intelligence is not a simple combination of embedded systems and AI. Its scope is broad and sophisticated, covering machine learning algorithms, embedded hardware, sensing systems, communication, runtime systems, memory management, energy optimization, real-time computing, security, privacy, and system reliability. Currently, there is no single formal and universally acknowledged definition of Embedded Artificial Intelligence. Some researchers define Embedded AI broadly as a spectrum ranging from cloud-device collaborative inference to fully on-device intelligence. In this view, embedded intelligence can span multiple levels, from cloud-assisted execution to fully autonomous on-device learning and inference.

  • In our lab, we study this full spectrum of Embedded Artificial Intelligence, from cloud-device collaboration to fully on-device intelligence. A major focus of our research has been all-on-device machine learning on embedded systems, especially for deep neural networks. We call this direction Embedded Deep Intelligence: enabling deep learning models to run, adapt, and learn directly on resource-constrained embedded platforms.

  • Embedded Deep Intelligence addresses a fundamental challenge: modern deep learning models are often designed under the assumption of abundant memory, computation, energy, and connectivity, while embedded systems operate under severe constraints. Our research investigates how to redesign deep learning algorithms and systems so that they can be executed and trained on real devices with limited memory, limited energy, limited computation, and strict latency requirements.

• From Embedded AI to Embedded Physical AI 

  • While Embedded Deep Intelligence focuses on enabling deep learning on embedded systems, Embedded Physical AI expands the scope from on-device inference and learning to closed-loop intelligence in the physical world.

  • Traditional embedded AI systems often follow a simple pipeline:

    • sensor input → on-device model → prediction 

  • Embedded Physical AI requires a richer and more dynamic pipeline: 

    • sensing → perception → memory → learning → reasoning → planning → action → feedback → adaptation 

  • In other words, Embedded Physical AI systems must not only classify sensor data, but also understand context, remember past observations, adapt to changing environments, make decisions under uncertainty, interact with actuators or users, and continuously improve over time. Examples include home robots that adapt to individual households, vehicles that process sensor data under real-time constraints, wearable devices that personalize models to users, smart environments that coordinate distributed sensors, and industrial systems that detect, reason about, and respond to physical events. 

  • The key research question is: 

    • How can AI systems continuously learn, reason, and act on real-world embedded devices under strict resource constraints? 

• Benefits of Embedded (On-Device) Physical AI 

  • Cloud-based machine learning solutions were widely used when wireless sensor networks mainly collected data from sensor nodes and transmitted them to remote servers for later analysis. However, today’s embedded systems are far more capable in terms of processing power, memory, sensing capability, and energy efficiency. Modern microcontrollers, mobile processors, edge accelerators, and neural processing units can execute increasingly sophisticated machine learning workloads locally.

  • On-device machine learning and Embedded Physical AI offer several important advantages over cloud-only approaches.

    • Data Transmission Cost and Latency. Data communication between a device and a cloud server or base station introduces latency and energy overhead. In physical-world applications, delayed responses can lead to degraded user experience or even unsafe behavior. A robot cannot always wait for a cloud response before moving; a wearable device cannot continuously transmit sensitive physiological signals; a vehicle must process sensor data within strict deadlines. Embedded Physical AI reduces communication dependency by processing data close to where it is generated and where action must be taken. 

    • Privacy and Security. Many embedded devices collect private and sensitive data, including health signals, audio, video, location, behavior patterns, and personal context. Processing such data locally can reduce exposure to external entities and lower the risk of privacy leakage. Embedded Physical AI enables privacy-preserving intelligence by keeping sensitive data on the device whenever possible, while still allowing the system to learn and adapt to the user or environment. 

    • Precision Learning and Resource Management. Many human-centered AI applications require personalization. Wearable, implantable, mobile, robotic, and smart-home systems often interact with different users who have different behaviors, preferences, physical conditions, and expectations. On-device learning allows these systems to adapt to individual users and local environments. This enables more precise, personalized, and context-aware intelligence than one-size-fits-all cloud models. 

    • Adaptability and Lifelong Learning. Physical environments are dynamic. Lighting changes, sensors degrade, users behave differently over time, objects move, machines age, and deployment conditions shift. Embedded Physical AI enables devices to adapt throughout their lifetime by learning from local data and feedback. This is particularly important for robotics, autonomous systems, smart environments, and long-term sensing applications, where the system must remain useful after deployment rather than relying only on offline training. 

    • Real-Time and Reliable Physical Interaction . Unlike purely digital AI systems, Embedded Physical AI systems interact with the physical world. They must respond under real-time constraints, handle uncertainty, recover from failures, and operate reliably over long periods. This requires system-level support for adaptive computation, deadline-aware execution, uncertainty monitoring, fallback mechanisms, and robust sensing. Embedded Physical AI therefore shifts the goal from simply achieving high model accuracy to achieving deployable and reliable intelligence in real-world conditions. 

Why do we research Embedded AI and Embedded Physical AI ?

• Embedded systems are everywhere. 

• Embedded systems are the dominant computing platform of today. They are deeply integrated into the physical world and continuously interact with people, environments, machines, and infrastructure. 

• Billions of embedded, mobile, wearable, robotic, vehicular, and IoT devices are deployed worldwide, generating massive amounts of data from the real world. These devices include smartphones, smartwatches, IoT sensors, robots, drones, autonomous vehicles, smart appliances, medical devices, industrial machines, and cyber-physical systems. They far outnumber traditional general-purpose computing platforms such as desktops, laptops, workstations, and servers, and they cover an enormous range of real-world applications. 

• Every time we look at a smartwatch, answer a phone, take a picture, drive a car, turn on a TV, use a home appliance, or interact with a robot, we are interacting with an embedded system. These systems are not distant computing resources hidden in data centers; they are the computing platforms closest to our daily lives and the physical world around us. 

• This ubiquity makes embedded systems the natural foundation for the next generation of artificial intelligence. 

• Embedded AI brings artificial intelligence from data centers to everyday life. 

  • Today’s state-of-the-art artificial intelligence often requires enormous computing resources, including high-end GPUs, large memory, large-scale storage, and substantial energy. Such resources are typically available only in large data centers, major corporations, or well-funded research laboratories. As a result, many powerful AI technologies remain concentrated in centralized computing infrastructures. 

  • In contrast, most people interact with AI through small and resource-constrained devices, such as smartphones, wearables, sensors, home appliances, and mobile robots. These devices have limited processors, small memory, constrained battery capacity, intermittent connectivity, and strict latency requirements. They cannot simply run today’s large-scale AI models in the same way cloud servers do. 

  • Embedded Artificial Intelligence, or Embedded AI, addresses this gap. It aims to enable AI directly on or near embedded devices, allowing them to process data locally, make decisions quickly, protect sensitive information, and adapt to users and environments. By bringing AI to the enormous number of embedded systems around us, Embedded AI can democratize artificial intelligence and make its benefits available through everyday devices rather than only through centralized cloud infrastructure. 

  • Embedded AI is therefore not merely about shrinking AI models. It is about redesigning AI algorithms, system architectures, memory management, runtime execution, and hardware-aware optimization so that intelligence can operate under real-world resource constraints. 

• Embedded Physical AI extends Embedded AI from prediction to action. 

  • While Embedded AI focuses on enabling intelligence on resource-constrained devices, Embedded Physical AI extends this vision further. Embedded Physical AI systems do not only analyze data; they sense, learn, reason, plan, adapt, and act in the physical world. 

  • Traditional embedded AI systems often follow a simple pipeline:

    • sensor input → on-device model → prediction

  • Embedded Physical AI requires a richer closed-loop pipeline:

    • sensing → perception → memory → learning → reasoning → planning → action → feedback → adaptation

  • This shift is important because physical-world intelligence is fundamentally different from purely digital intelligence. A physical AI system must respond under real-time constraints, handle uncertainty, recover from failures, adapt to changing environments, and operate reliably over long periods. It must also satisfy strict memory, latency, energy, reliability, and hardware constraints.

  • For example, a robot cannot always wait for a cloud server before taking action. A wearable device cannot continuously transmit sensitive health data to a remote data center. An autonomous vehicle must process sensor data and make decisions within strict deadlines. A smart home system must adapt to different users, environments, and privacy requirements. These applications require AI that is embedded, adaptive, and physically grounded.

  • Embedded Physical AI therefore shifts the goal from simply achieving high model accuracy to building AI systems that are deployable, adaptive, reliable, and capable of real-world physical interaction.

• Embedded AI and Embedded Physical AI enable local, private, and adaptive intelligence. 

  • The data generated by embedded devices is enormous, diverse, and deeply connected to real life. Processing this data locally creates new opportunities to improve everyday life without relying entirely on distant data centers.

  • On-device and embedded intelligence can address several major limitations of cloud-centric AI:

  • First, it can reduce communication cost and latency by processing data close to where it is generated. This is crucial for real-time applications such as robotics, vehicles, wearables, industrial monitoring, and interactive sensing systems.

  • Second, it can improve privacy and security by keeping sensitive data on local devices. Personal health signals, audio, video, location traces, and behavioral patterns do not always need to leave the device to be useful.

  • Third, it can enable personalization and precision learning. Different users, environments, machines, and physical contexts require different AI behaviors. Embedded AI allows systems to adapt locally to each user and deployment condition.

  • Fourth, it can reduce energy consumption and cloud dependence by avoiding unnecessary data transmission and centralized computation.

  • Finally, it can support lifelong adaptation. Physical environments change over time, and intelligent devices must continue to learn after deployment. Embedded Physical AI enables devices to evolve with their users and environments rather than remaining fixed after offline training.

• Embedded AI will facilitate collaborative and collective intelligence at scale. 

  • The long-term vision of Embedded AI and Embedded Physical AI is not limited to individual intelligent devices. Once billions of embedded devices become capable of local learning, adaptation, and decision-making, they can also collaborate with one another to form large-scale collective intelligence.

  • In the first phase, Embedded AI enables individual devices to learn, evolve, and adapt locally without relying entirely on external systems. Each device becomes more capable of understanding its own user, environment, and operating conditions.

  • In the second phase, connected embedded devices can share knowledge, models, summaries, or representations with one another. Instead of transmitting all raw data to the cloud, devices can exchange useful learned information while preserving privacy, reducing communication cost, and improving system-wide intelligence. This enables collaborative sensing, distributed learning, multi-device inference, and edge-cloud intelligence.

  • In the third phase, networks of embedded and physical AI systems can autonomously coordinate, learn, and act together. Robots, vehicles, sensors, wearables, smart appliances, and edge servers may collectively observe the physical world, exchange knowledge, and make better decisions than any single device could make alone. This will enable a new form of distributed physical intelligence that continuously improves through interaction with the real world.

What is the goal of our lab?

• • We pursue excellence in research on electrical engineering, computer science, and artificial intelligence.

    • We relentlessly push forward state-of-the-art science and technology via disruptive ideas and game-changing perspectives. 

    • We imaginatively turn our ideas into meaningful artifacts (e.g., software, hardware, paper, etc.) that can solve the real problem in the world, which makes our research practical and useful.

    • We actively communicate with peer researchers worldwide for discussion, facilitating the progress of research in the community.

• • We make the world a better place by making real impacts with our research.

    • We contemplate how our research can improve the lives of people; not only academically but also socially and practically.

    • We help people solve people's problems by making intellectual contributions to society via our original, innovative, novel, and sound ideas. 

    • We do our best to explain and publicize our work to the public so that they can correctly appreciate and utilize our research as they need. 

• • We collaborate with and learn from each other when solving challenging problems.

    • We enjoy discussing not only with our lab members having similar ideas and thoughts but also other people out of the field who can introduce fresh ideas and points of view. 

    • We help each other as a team member since one person does not have all the necessary ability to conduct successful research.

    • We are unafraid of conducting interdisciplinary work with people from various backgrounds via consilience and convergence of the studies.