SEOUL, July 06 (AJP) - Artificial intelligence that can see through a window, judge how light bounces off a surface, and think a step ahead before acting, arrived this week from a Korea Advanced Institute of Science and Technology research team. The four new technologies, unveiled together, push AI systems beyond recognizing what is in an image and toward understanding and reacting to the physical world the way a person does, a shift researchers call "physical AI."
The work matters because it targets a problem that has held back self-driving cars, humanoid robots, and other machines meant to operate in real spaces. Those systems still struggle with basic physical realities like glass, changing light, and unfamiliar surroundings. The four papers were accepted at two of the field's most competitive conferences, the International Conference on Learning Representations (ICLR) 2026 and the Conference on Computer Vision and Pattern Recognition (CVPR) 2026, two as oral presentations and two as highlight papers, a distinction reserved for standout research. KAIST said the oral presentations ranked among the top 1.13 percent of ICLR submissions and the top 0.8 percent of CVPR submissions.
The Korea Advanced Institute of Science and Technology (KAIST)'s Yoon Sung-eui, a professor in the School of Computing, led the research. KAIST said Monday that the four projects connect into a single technical chain, an AI system that sees, understands what it is looking at, predicts what happens next, and plans its own actions accordingly.
The first technology, called GLINT, tackles a problem that has quietly frustrated computer vision for years: glass. People instinctively separate what they see reflected in a window from what lies beyond it. Most AI systems cannot make that distinction, and they process both as a single, confusing image. Yoon's team built GLINT to split the reflection from the transmitted scene, allowing the AI to correctly interpret spaces filled with windows, mirrors, or other transparent surfaces. The paper, "GLINT: Modeling Scene-Scale Transparency via Gaussian Radiance Transport," was presented orally at CVPR 2026 on June 5.
The second technology, RadioGS, focuses on light itself. The same object looks different in sunlight than it does under indoor lighting, and that variation has long thrown off AI vision systems. RadioGS trains an AI to understand how light reflects off and scatters across a surface, so it can recognize an object's true material and surroundings even as the surrounding lighting changes. The paper, "Radiometrically Consistent Gaussian Surfels for Inverse Rendering," was presented orally at ICLR 2026 on April 23.
The third technology, called Visual-RRT, gives robots a way to navigate using only a photograph. Conventional robots need exact coordinates to find their way somewhere. Visual-RRT instead lets a robot compare what it currently sees with a single target image and figure out its own path there. In physical tests, the robot reached its destination successfully using nothing more than that one photo, a result the team said could apply to service robots and self-driving delivery robots alike. The paper, "Visual-RRT: Finding Paths toward Visual-Goals via Differentiable Rendering," was selected as a CVPR 2026 highlight paper on March 27.
The fourth technology, CLaD, addresses planning. Before people act, they tend to think a step ahead, asking what will happen if they move a certain way. CLaD gives AI systems the same habit, having them forecast the likely outcome of an action before choosing the best one to take. KAIST said the approach lets AI complete tasks with a higher success rate even in complex environments, which the team expects to make it a core technology for the next generation of autonomous robots. The paper, "CLaD: Planning with Grounded Foresight via Cross-Modal Latent Dynamics," was also selected as a CVPR 2026 highlight paper.
Yoon said the four results mark a shift in what AI is capable of. "This research shows that AI is evolving beyond simply seeing with its eyes, toward understanding the real world and predicting what will happen next in order to decide its own actions," he said. "We hope this achievement contributes to the advancement of physical AI technologies that operate in real environments, including self-driving cars and humanoid robots."
(Reference Information)\
Journal/Source: Conference on Computer Vision and Pattern Recognition 2026
Title: GLINT: Modeling Scene-Scale Transparency via Gaussian Radiance Transport
Link/DOI: https://arxiv.org/abs/2603.26181
Journal/Source: International Conference on Learning Representations 2026
Title: Radiometrically Consistent Gaussian Surfels for Inverse Rendering
Link/DOI: https://arxiv.org/abs/2603.01491
Journal/Source: Conference on Computer Vision and Pattern Recognition 2026
Title: Visual-RRT: Finding Paths toward Visual-Goals via Differentiable Rendering
Link/DOI: https://arxiv.org/abs/2604.16388
Journal/Source: Conference on Computer Vision and Pattern Recognition 2026
Title: CLaD: Planning with Grounded Foresight via Cross-Modal Latent Dynamics
Link/DOI: https://arxiv.org/abs/2603.29409
Copyright ⓒ Aju Press All rights reserved.