Research & Projects
Where intelligent systems meet real-world impact.
The CarAI Lab conducts applied and fundamental research at the intersection of artificial intelligence, machine learning, and autonomous systems. Our projects across three research areas address critical challenges in autonomous driving, adaptive robotics, and intelligent control, developing solutions that are both technically rigorous and practically deployable.
Research Areas & Projects
Occupancy Networks in Autonomous Driving
We are developing the next generation of 3D perception for autonomous vehicles through advanced Occupancy Networks. Moving beyond traditional 3D bounding boxes and 2D bird’s-eye views, our research focuses on dense 3D volumetric representations that jointly predict spatial occupancy and semantic labels in real time. This approach allows systems to capture fine-grained geometric details and accurately detect both familiar objects and unexpected, out-of-vocabulary obstacles.
Vision-Language Models (VLMs)
We are building VLMs to improve high-level scene understanding, reasoning, and decision-making in Autonomous Vehicles (AVs). With integration of multimodal AI architectures, our research enables self-driving systems to interpret complex semantic context, e.g., gestures from traffic control officers or written road signs, that traditional perception pipelines miss. This fusion of vision and linguistic reasoning improves vehicle explainability and allows AVs to navigate long-tail, ambiguous driving scenarios with human-like intuition and safety.
Vision-Language-Action (VLA) Models
We are advancing the state-of-the-art in embodied intelligence by building VLA models that directly bridge high-level perception and reasoning with low-level motion control. Unlike traditional pipelines that separate understanding from planning, our VLA models translate multimodal inputs, including real-time visual feeds and natural language commands or contextual goals, directly into continuous control signals and trajectory predictions. This end-to-end integration enables AVs and mobile robots to execute complex, multi-step actions and adapt to dynamic real-world environments through unified foundational intelligence.
Morphology-agnostic robotics
Our lab builds highly adaptable robotic systems that are not limited by their physical shape or structure. Using advanced VLMs, VLAs, and reinforcement learning, we design AI that enables robots to intelligently adapt to and recover from unexpected hardware failures, such as damaged joints or compromised sensor suites.
This research opens new possibilities for deploying robots in challenging real-world environments where hardware reliability cannot be guaranteed.