Researchers at the Gwangju Institute of Science and Technology (GIST) have developed an artificial intelligence-based robot navigation technology that can understand and locate objects in a three-dimensional space based on human language descriptions.
GIST said April 2 that a research team led by Professor Kim Eui-hwan from the Department of AI Convergence created the technology, called Context-Nav, which enables robots to interpret detailed sentences describing objects and accurately identify their locations.
The system analyzes not only physical characteristics such as color and shape but also the relative positions of objects, demonstrating potential for expansion into various service robot applications.
Previous approaches faced limitations in fully incorporating contextual elements from human instructions, such as spatial relationships with surrounding objects, relative direction and arrangement, and situational cues embedded in longer descriptions.
Spatial relationships — such as left, right, front and back — can vary depending on the observer’s perspective and position, increasing the likelihood that robots misidentify targets using conventional methods.
To address this, the research team developed a method that allows robots to use entire sentence descriptions during navigation, enabling a combined understanding of object features and three-dimensional spatial relationships with surrounding items.
For example, when given the instruction, “Find the red book on the table next to the living room sofa,” the robot interprets the sentence not just as object data but as positional information within a 3D space.
The technology also demonstrated strong performance in CoIN-Bench, a benchmark designed to evaluate a robot’s ability to understand long sentences containing relationships between objects as well as detailed attributes such as color and shape.
Unlike conventional reinforcement learning methods, where robots learn optimal actions through repeated trial and error, the new approach achieved a success rate of 20.3% without additional training. This is approximately 2.3 times higher than traditional reinforcement learning-based methods, which recorded a success rate of 8.9%.
Professor Kim said the research can be applied immediately to new environments or unfamiliar objects without requiring additional task-specific training or adjustments, adding that it could serve as a key enabling technology for real-world applications of indoor service robots and intelligent robotic systems.
The study, conducted under Kim’s supervision and carried out by Jang Won-sik, was published in advance on the international academic server arXiv on March 18.
Discussions regarding technology transfer can be arranged through the institute’s technology commercialization office.
