Professor Hyun Myung’s research team (Urban Robotics Lab) from our department secured second place in the NaviTrace Challenge, held during the Open-World Navigation (OWN) Workshop at RSS 2026 (Robotics: Science and Systems) in Sydney, Australia, from July 13 to 17. The international competition evaluated an AI’s ability to infer navigation paths on a given image based solely on a single first-person perspective photograph and a brief natural language instruction.
Commanding a robot to “cross the street” provides no explicit information about finding a crosswalk or waiting for a signal; the robot must independently decipher these implicit social norms from the scene before it. However, assigning goal identification, hazard assessment, and path generation to a single AI model leads to task interference and degraded performance. To address this, the research team developed PRISM-Nav, which divides the process among four specialized agents. Three agents simultaneously identify target points, hazardous elements such as stairs or curbs, and socially relevant structures like crosswalks or sidewalks, while a final agent synthesizes their outputs to generate the path. Crucially, the agents communicate by drawing symbols directly onto the input photograph rather than exchanging text or coordinates, significantly reducing spatial information loss—much like marking a location on a map instead of describing it verbally. Because it requires no additional training, PRISM-Nav can be immediately deployed to new robots or environments.
The team scored 53 points, placing second behind the joint team from Nanjing University and FiveAges. A key highlight of the achievement was the economic efficiency of the underlying AI models. While the top-performing general-purpose model provided as a reference by the organizers (Gemini 3.1 Pro Preview) carries high operational costs that make continuous deployment on real robots difficult, the team achieved performance surpassing that benchmark using a far more affordable lightweight model (Gemini 3 Flash). This demonstrates that deploying multiple low-cost agents offers a clear advantage in both cost and performance over relying on a single expensive model.
Professor Myung noted, “This result proves that large AI models can be effectively utilized for real-world decision-making in robotics without requiring additional training data. It will serve as a foundational technology for service robots, such as delivery and guide robots, that share physical spaces with humans.”
Meanwhile, Professor Myung’s research group previously secured first place in international challenges at the ICRA 2026 and CVPR 2026 workshops last June. The laboratory continues to expand its expertise in spatial perception toward the field of Embodied AI.
