"Finding with Human Descriptions"... GIST Develops AI Robot Navigation Technology
(From left) Professor Euihwan Kim of the AI Convergence Department, Jang Wonsik, Integrated Master's and Doctoral Course Student.
View original imageGwangju Institute of Science and Technology (GIST) announced on April 2 that a research team led by Professor Kim Euihwan from the Department of AI Convergence has developed an artificial intelligence (AI) robot navigation technology called "Context-Nav." This technology allows a robot to understand and accurately locate objects described by humans in sentences within a 3D space.
This technology analyzes not only the physical characteristics of objects, such as color and shape, but also their relative positions to other items, demonstrating its potential for expansion into various service robot applications.
For autonomous robots (service robots) that perform a variety of tasks in indoor environments—such as cleaning, delivery, and guidance—to understand and accurately execute human language instructions, they must comprehensively grasp the spatial relationships between objects around them.
In previous studies, the "reinforcement learning" approach was mainly used, where robots repeatedly attempt and fail to find the optimal action strategy on their own. However, this method requires a vast amount of data and incurs high training costs.
Three-Dimensional Context-Based Robot Navigation Process Using Long Sentence Descriptions.
View original imageTo address these issues, the research team proposed a method in which the entire long sentence description provided by a human is utilized during the robot's navigation process. This enables the robot to simultaneously understand the target object's characteristics and the 3D spatial relationships among surrounding objects. For example, if a person says "Find the red book on the table next to the sofa in the living room," the robot interprets the sentence not as simple object information but as positional information within a 3D space.
First, the robot uses an RGB camera and a depth sensor to recognize its surrounding environment, and then it identifies in real time the areas most likely to match the description. It calculates how well each candidate area fits the target and records the suitability score on a "Value Map."
Subsequently, the robot determines an efficient search path centered on the area with the highest score. When a candidate object is found, it uses a vision-language model—which comprehends both image and text information—to verify object attributes. It then precisely verifies the spatial relationships with surrounding objects (above, below, left, right, front, back) through 3D spatial reasoning.
The research team's technology also demonstrated high performance in the "CoIN?Bench" target-finding test, which evaluates the robot's understanding of long sentences containing detailed attributes such as relationships with objects, color, and shape. Unlike conventional reinforcement learning—where the robot repeatedly learns optimal actions through trial and error—the team's technology achieved a 20.3% success rate without additional training, which is about 2.3 times higher than the previous reinforcement learning-based method (8.9%).
In particular, the strategy of interpreting human descriptions in 3D space, reflecting the entire sentence context to begin navigation in the most likely locations, and verifying spatial relationships with surrounding objects has proven highly effective in improving the accuracy of the robot's actions. Furthermore, the more the entire long sentence is reflected in the navigation process, the less unnecessary movement occurs, and the process of confirming spatial relationships in 3D from multiple perspectives also reduces cases of misidentifying the target.
Professor Kim Euihwan stated, "This research presents a precise technology that enables robots to understand not only the object's own features but also its contextual and 3D spatial relationships. Because it can be applied immediately to new spaces or unfamiliar objects without task-specific training or adjustments, it will become a core foundational technology for enhancing the practical usability of indoor service robots and intelligent robotic systems in the future."
Hot Picks Today
"Two Spend 500,000 Won in a Flash, Yet Wait Two Hours?" ... 'Dreamland' Shunned by Kids and Foreign Tourists
- The Era of Getting 200 Million Won After 30 Years Is Over... A Japanese Bank's "Perpetual Department Head" Who Just Clocked In Cries Out
- Spain Hit by Its Worst Tragedy... Lee Kangin, Moreno and Real Madrid Join Soccer World in Mourning
- They Said, “Make It Your Home for Life” - Then Suddenly Kicked 200 Passengers Off... What Happened to the Around-the-World Cruise?
- "130 Rice Bowls Sold Out in a Day"... 'Unfiltered Face' Adds 8 Million Won to Night Market Sales: China's Bungeunwoo
This research, supervised by Professor Kim Euihwan and conducted by Jang Wonsik, an integrated master's and doctoral course student at the Department of AI Convergence, was supported by the Excellent Young Researcher Program of the Ministry of Science and ICT and the National Research Foundation of Korea, the IITP’s Problem Hypothesis and Self-Supervised Visual Intelligence Technology Development Project, and the NST Global TOP Strategic Research Group Project. The research results were pre-published on the international academic server "arXiv" on March 18, 2026, and are scheduled to be presented at the prestigious international AI conference, the Computer Vision and Pattern Recognition Conference (CVPR 2026).
© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.