From left: Park Dae-hee, a professor in the Department of Electrical Engineering and Computer Science at DGIST, master's student Seo Tae-won, and undergraduate researcher Jeon Seon-ae. [DGIST]
From left: Park Dae-hee, a professor in the Department of Electrical Engineering and Computer Science at DGIST, master's student Seo Tae-won, and undergraduate researcher Jeon Seon-ae. [DGIST]

DGIST announced Saturday that a research team led by Professor Park Dae-hee of its Department of Electrical Engineering and Computer Science, working jointly with a KAIST research team, has developed a learning technology that enables a single compact AI model to simultaneously predict the movements of nearby people and plan a robot's safe navigation path — while minimizing performance degradation across both tasks.

The findings were accepted by ECCV 2026, one of the world's top three international conferences in computer vision, and were presented at the conference held in Malmö, Sweden, from Tuesday through Saturday.

Robots navigating crowded spaces must predict how nearby people will move while at the same time planning a safe route that avoids collisions.

Running the two tasks on separate AI models demands more computing resources and memory, making deployment on actual robots impractical — which is why a technology that lets a single compact AI model handle both tasks together is needed.

To address this, the research team developed what they call "Disjoint Parameter Training" (DPT), a method that trains each function to rely on distinct internal components of the AI.

The team then selectively merged only the parameters essential to each function into a single model, allowing both capabilities to operate without interfering with each other while keeping the model small.

Performance validation using benchmark datasets in the robot navigation field — including JRDB and JTA — showed that the approach predicted the movements of nearby people more accurately than existing methods and reduced both navigation path error and the likelihood of collisions.

Additional experiments applying the technology to autonomous driving AI also showed performance gains, confirming its potential for use across a range of physical AI applications including robotics and autonomous driving.

"This research addresses the performance degradation that occurs when multiple functions are packed into a single small AI," Park said. "We plan to apply it to actual robots to verify its practical utility."

The study was co-authored by DGIST master's student Seo Tae-won and undergraduate researcher Jeon Seon-ae, with Park serving as corresponding author. The paper, titled "Unified Prediction and Planning via Conflict-Aware Disjoint Parameter Training," was supported by the NVIDIA Academic Hardware Grant Program.


kbj7653@heraldcorp.com