Research Hub: (AI Research Hub)
Current Research Interests:
----Physical AI/ Embodied Intelligence
----Machine Learning / Deep Learning / Large Language Models(LLMs)
----Bayesian Network
----Real-Time Computing(RTC)
----Ubiquitous/Pervasive Computing
----Mobile and Location-aware Computing
Embodied Intelligence is a branch of artificial intelligence where machines possess physical or virtual bodies and learn to navigate the world through real-time interaction. Instead of relying solely on pre-programmed instructions or offline data, these systems integrate perception, continuous sensorimotor feedback, and action to adapt to unpredictable environments.
How Embodied Intelligence Works
At its core, Embodied Intelligence requires the synergy of three fundamental elements:
1. The Brain (Intelligence): Advanced AI algorithms, like Large-Scale Foundation Models, that process high-complexity inputs and enable reasoning, planning, and decision-making.
2. The Body (Embodiment): The physical hardware or simulated morphology (such as a humanoid robot, robotic arm, or autonomous vehicle) that executes movements and gathers physical data through sensors.
3. The Environment: The external, unpredictable real world in which the agent operates and interacts.
By tightly closing the "perception-action loop," the system senses the environment, decides on a course of action, acts, and then observes the outcome, constantly refining its behavior and internal models.
Key Applications
1. Industrial Manufacturing: General-purpose and humanoid robots (such as UBTech's Walker S1) are being deployed on assembly lines to adjust posture, grip, and balance dynamically while working alongside human employees.
2. Logistics & Warehouses: Autonomous Mobile Robots (AMRs) orchestrate complex inventory sorting and transport, relying on flexible software to dynamically navigate unstructured, fast-changing facilities.
3. Healthcare: Embodied devices are being developed to assist in complex robotic surgeries, lung cancer diagnostics, and emergency evacuations.
4. Autonomous Driving: Self-driving cars continuously perceive obstacles, predict human dynamics, and physically maneuver through varying traffic and weather.
Why Embodied Intelligence Matters
While traditional AI remains constrained to digital environments, embodied intelligence anchors cognition in the physical world. By facing the physical limitations of real-world interactions (e.g., gravity, friction, and uncertain physics), agents learn to generalize skills across open-ended scenarios. Many researchers view this adaptive, cross-domain learning as a crucial pathway toward achieving Artificial General Intelligence (AGI).
Physical AI integrates artificial intelligence with hardware—such as sensors and actuators—to enable machines, robots, and autonomous vehicles to perceive, understand, reason, and act within the physical world in real-time. Unlike digital-only AI, it bridges the gap between digital intelligence and physical interaction, allowing systems to adapt to dynamic environments.
How Physical AI Works
Physical AI systems use multimodal inputs, including cameras, LiDAR, and sensors, to process data and make decisions.
1. Perception & Modeling: AI understands spatial relationships and physical rules of the 3D world, moving beyond static, pre-programmed instructions.
2. Actionable Intelligence: Algorithms (including reinforcement and simulation learning) enable robots to perform tasks such as grasping, navigating, or manipulating objects, says Physical Intelligence.
Benefits and Key Applications
Physical AI is transforming industries by automating complex physical labor, reducing costs, and boosting productivity.
1. Manufacturing: Adaptive assembly lines and robotic, self-driving forklifts, says The World Economic Forum.
2. Transportation: Safer self-driving cars and autonomous vehicles.
3. Healthcare: Robotic assistants for surgery or patient care.
4. Smart Systems: Robotic agents and intelligent infrastructure for smart grids.
Large Language Models(LLMs) are deep learning AI systems trained on massive text datasets to understand, generate, and summarize content, often using transformer architectures with self-attention. Working by predicting the next word in a sequence, LLMs are key in automation, chatbots, translation, and code generation.
How Large Language Models Work
1. Training Method: LLMs learn from vast datasets, analyzing word relationships to predict the next word or phrase in a sequence.
2. Transformer Architecture: The underlying neural network utilizes self-attention to process text, identifying connections between words regardless of distance.
3. Parameters: They are characterized by a massive number of parameters (memory) that enable sophisticated reasoning and content generation.
Components of LLMs
1. Pre-training: Initial phase where the model learns language patterns from massive, unlabeled text data.
2. Fine-tuning: Specialized training on a smaller, labeled dataset to excel at specific tasks like customer service or medical analysis.
3. Encoder/Decoder: Neural network layers that extract meaning from input sequences (encoder) and generate output text (decoder).
Key Features and Types
1. Versatility: Capable of creative writing, coding, language translation, and summarization.
2. Multimodality: Modern LLMs, or "foundation models", are increasingly working across text, image, and audio formats.
3. Types: Examples include proprietary models like ChatGPT (OpenAI), Gemini (Google), and Claude (Anthropic), alongside open-source alternatives like Llama (meta).
Applications of LLMs
1. Business: Customer support chatbots, automated email generation, and document summarization.
2. Healthcare: Analyzing research papers, protein structure prediction, and electronic health record analysis.
3. Software Development: Writing, debugging, and documenting code.
4. Finance: Analyzing market trends and creating transcripts of meetings.
Bayesian network is a probabilistic graphical model that represents variables and their conditional dependencies using a Directed Acyclic Graph (DAG). It calculates the likelihood of events occurring by combining prior knowledge with new evidence, often used for reasoning under uncertainty in fields like medical diagnosis, AI, and anomaly detection.
Key Components
1. Nodes: Represent random variables (e.g., "Disease," "Symptom").
2. Edges: Directed arrows showing direct dependencies between nodes.
3. Conditional Probability Tables (CPT): Each node contains a CPT, quantifying the effect of parent nodes on the child node.
How a Bayesian Network Works
1. Structure Definition: Variables are linked in a directed graph where edges show dependencies.
2. Probability Assignment: Nodes are filled with CPDs (Conditional Probability Distributions), often using Bayes' theorem to update beliefs as new evidence is observed.
3. Inference: Using algorithms, the network calculates the probability of specific outcomes (e.g., predicting the likelihood of a disease given symptoms).
Types of Bayesian Networks
1. Static Bayesian Network: Standard model representing a single point in time.
2. Dynamic Bayesian Network (DBN): Models variables over adjacent time steps (e.g., Hidden Markov Model).
Key Advantages
1. Handling Uncertainty: Explicitly models uncertainty using probability.
2. Causal Insight: Can represent causal relationships rather than just correlations.
3. Domain Knowledge Integration: Can be constructed by experts or learned from data, or a combination of both.
Applications
1. Medical Diagnostics: Modeling diseases and symptoms.
2. Industrial Fault Diagnosis: Root cause analysis for machinery.
3. Risk Assessment: Determining probabilities of insurance claims or hazards.
4. Machine Learning: Anomaly detection and decision-making.
Real-time computing refers to hardware and software systems that process data and respond to events within strict, guaranteed time constraints or deadlines. These systems ensure reliability by processing information almost instantaneously—often within microseconds—crucial for applications like, air bag deployment, industrial robotics, and patient monitoring.
Key Aspects and Functions
1. Predictable Timing: Unlike fast systems, real-time systems are not just fast, but predictable. Failure to meet deadlines results in system failure.
2. Hard vs. Soft Real-Time: Hard real-time systems (e.g., flight control) must meet deadlines perfectly, as missing one is catastrophic. Soft real-time systems (e.g., streaming media) have deadlines, but missing them reduces quality rather than causing total failure.
3. Components: Components include specialized sensors/actuators (input), a real-time operating system (RTOS), and fast processors (e.g., VxWorks, QNX, and FreeRTOS from Cranes Varsity).
Benefits and Applications
1. Safety and Reliability: Essential for automotive, medical, and aerospace systems.
2. Automation: Powers AI defect detection on production lines.
3. Instant Interaction: Enables high-frequency trading, IoT, and real-time gaming.
Common Examples
1. Automotive: Airbag controllers, anti-lock braking systems (ABS), and adaptive cruise control.
2. Industrial: Robotic arms in manufacturing, PLC (Programmable Logic Controllers).
3. Medical: Pacemakers and bedside patient monitors.
4. Consumer: Gaming consoles, video conferencing, and live video streaming.
Disadvantages
1. Complexity: Difficult to design, implement, and test.
2. Cost: High expense for specialized hardware and software.
3. Low Flexibility: Difficult to modify once developed.
Ubiquitous/Persavasive Computing, also named as Ubi-Comp, is a a very interesting and hot research field. The term of "ubiquitous computing" was coined by Dr. Mark Weiser in 1988 which means the processing of any data at any time and anywhere through any network using any device.
5 As in ubiquitous computing: Any time, Any where, Any data, Any network and Any device.
"Ubiquitous computing names the third wave in computing, just now beginning. First were mainframes, each shared by lots of people. Now we are in the personal computing era, person and machine staring uneasily at each other across the desktop. Next comes ubiquitous computing, or the age of calm technology, when technology recedes into the background of our lives."------Mark Weiser
Ubiquitous Computing Vs Virtual Reality
Ubiquitous computing is roughly the opposite of virtual reality. Where virtual reality puts people inside a computer-generated world, ubiquitous computing forces the computer to live out here in the world with people. Virtual reality is primarily a horse power problem; ubiquitous computing is a very difficult integration of human factors, computer science, engineering, and social sciences.
Mobile Positioning Vs Mobile Location
Mobile positioning and mobile location are sometimes used interchangeably in conversation, but they are really two different things.
----Mobile positioning refers to determining the position of the mobile device.
----Mobile location refers to the location estimate derived from the mobile positioning operation.
Mobile Location Estimation System(MLES)
What is LBS(Location Based Services)
Location Based Services, or LBS, are services provided to clients based on the geographical location information of the mobile device.
What is MLES?
MLES is the acronym of Mobile Location Estimation System. Its mission is to develop an advanced mobile location estimation system based on the existing mobile phone infrastructure. MLES has achieved an average location error of 65 meters in selected areas in Hong Kong. MLES differs from other location systems in that instead of using a single method for positioning, it consists of a brunch of location algorithms. Each algorithm is designed to solve a specific problem domain. MLES uses signal strength for positioning.
Mobile Location Estimation System(MLES) Applications
Elderly Care
According to the HK Police, there are about 700 elderly reported to be missing due to Parkinson disease and memory fading. Since only 1/3 of the cases are reported, the HK Police are estimating over 2000 elderly are missing every year. With the location based service, and with the mobility of the HK police force, the "missing" person should easily be found.
Blind Assistance
According to the HK Blind Union, there are 18,000 blind people in Hong Kong. And getting lost or getting disoriented in the middle of the city becomes a common problem. With an MELS enabled application, these people will know and be known where they are and commune within the city.
Education
An MLES application can be an excellent babysitter to students, especially to kindergarten and primary school students. It can tell you if a school bus is approaching so you can bring your kid to take the bus at best timing, avoiding long time waiting or even missing it. It can also remind you when your kid has arrived at school, or inform you when the kid leaves school earlier.
Tourism
MLES could give tourists a new experience while traveling. How nice it would be when tourists can get the background information of an attraction in their own language when they approach it; and how helpful it would be when they could receive information such as recommendation on where to shop and dine nearby. MLES applications act as a guide to help them explore place more comfortably.
Logistics
Information such as current traffic condition, current location and situation of a vehicle in a fleet etc is crucial to an efficient fleet and resource management of logistic companies. With location based service enabled, this real-time information can be captured by a control centre. Logistic companies then can make use of the information gathered to plan more cost-effective routes of the fleet to serve their customer better. MLES can also provide vehicles of a fleet a better security. An alert system and a tracking system can be installed to the vehicles so that owners the vehicles can be informed by the alert system in case their vehicles are stolen. They can then use the tracking system to locate the stolen vehicles.