City University of Hong Kong × Artificial Intelligence Research Center Advancing Cross-Domain AI Reasoning and Learning

As artificial intelligence moves beyond the laboratory to the unpredictable realities of autonomous driving and flexible manufacturing, it no longer faces standardized datasets. In the real world, AI must navigate an environment defined by sensory noise, extreme weather, and unpredictable behavior. To overcome these challenges, the City University of Hong Kong (CityU) and the AI Research Center have engaged in deep collaboration to develop next-generation agents capable of uncertainty reasoning and low-energy continual learning.

Equipping AI with Defensive Driving Logic

A core breakthrough of this collaboration lies in the field of uncertainty reasoning. Conventional AI models, when faced with adverse weather or sensor interference, often produce seemingly confident yet incorrect predictions that can lead to safety incidents. Our objective is to enable AI to adopt defensive driving logic similar to that of human drivers.

To achieve this, our team has integrated explicit uncertainty modeling at the perception layer. By dynamically weighting inputs from various sensors such as LiDAR and cameras, the system can autonomously determine which sensor is more reliable under extreme weather or during sensing failure. At the decision-making layer, the system no longer forces precise positional outputs but instead generates probability distributions with variance. As noise increases, the AI deliberately expands prediction variance, just as human drivers instinctively slow down and increase following distance in poor visibility. Through this mechanism, the agent tightens safety margins in the face of chaotic data, realizing decision logic closer to that of humans. Yet for AI to truly advance toward autonomy, it must also achieve continual learning.

Traditional models typically suffer from catastrophic forgetting, losing prior knowledge (such as basic traffic rules) when learning new skills (such as adapting to new road conditions). To address this challenge, the team has developed a growth strategy based on network positional importance. By adding new connections at highly important yet previously unlinked locations within the neural network, fresh knowledge can be stored without overwriting existing knowledge.

In addition, to reduce AI’s massive energy footprint, the research team implemented Spiking Neural Networks (SNNs), a next-generation architecture that computes only when signals are received and remains silent otherwise, fundamentally eliminating redundant operations. This means that future edge devices will possess an energy-efficient brain, awakening to think only when necessary and making real-time decisions at ultra-low power.

In addition, the team has developed ModeSeq technology, which elevates trajectory prediction from mere stochastic selection to sequential reasoning across multiple potential behaviors. This gives AI a step-by-step cognitive ability, enabling it to interpret pedestrian gestures, understand the meaning of a horn, and transform raw data into contextual scene anticipation.

Laying the Bedrock for General-Purpose Agents

Hon Hai Technology Group is currently making a full-scale push into electric vehicles (EVs) and robotics. Professor Jianping Wang of City University of Hong Kong notes that the ModeSeq system can achieve high-frequency reasoning on vehicles at low power consumption, truly moving AI from the laboratory to production lines and mass-market cars. In addition, this methodology will also benefit factory robots, transforming them from passive executors into safe, trustworthy, and intelligent collaborators.

Moving forward, the partnership aims to extend the research outcomes into industrial applications. The objective is to enable robots in complex warehouses or campuses to autonomously understand tasks, identify risks, and collaborate with humans.

Professor Wang concludes, “If general intelligence is a skyscraper, our current work represents the bedrock.” Through continual learning and robust reasoning, the collaboration between CityU and the AI Research Center is unlocking limitless possibilities for AI systems designed to operate long-term and evolve continuously.