KAIST Robot Systems Learn to Verify Destinations and Read Unwritten Rules

Researchers at the Korea Advanced Institute of Science and Technology have developed artificial intelligence systems designed to make robots more reliable when navigating unfamiliar environments. The work helped the university’s Urban Robotics Lab secure first and second place in two international robot navigation challenges, KAIST said.

The first system addresses a deceptively difficult problem: a robot may find a plausible route but incorrectly conclude that it has reached the requested destination. A joint team from KAIST and the Electronics and Telecommunications Research Institute created CoRe-VLN, or Coverage-based Recovery for Vision-Language Navigation, to reduce that type of error.

Before stopping, a robot using the system scans its surroundings again and compares what it sees with the original instruction. The AI examines details such as an object’s color, shape and proximity. If the scene does not match the target, the robot searches for another route and continues moving rather than ending the task prematurely. The researchers said the method can be added to existing navigation systems without additional training.

The joint team, competing as URL-FRRS, recorded an average success rate of 90.7 percent across two tasks in the VLNVerse Challenge, held in Malmo, Sweden, on Sept. 9 in connection with the European Conference on Computer Vision 2026. The result placed it first among 112 teams. A runner-up achieved the same success rate, but URL-FRRS took the top position because it submitted its results earlier. The third-place team recorded 89.7 percent.

KAIST’s second system focuses on rules that people often understand without being told, such as using a parking lot entrance instead of an exit or recognizing how sidewalks and crosswalks shape an appropriate route. Named PRISM-Nav, the system distributes the navigation problem among four AI agents. One identifies the destination, another detects hazards including stairs and curbs, a third interprets spatial and social rules, and a fourth combines their findings to choose a route.

Instead of exchanging lengthy descriptions or coordinates, the agents communicate by marking destinations and danger areas directly on an image. KAIST said the system can operate in new settings without retraining. It scored 53 points in the NaviTrace Challenge at the Open-World Navigation Workshop of Robotics: Science and Systems 2026 in Sydney, finishing second behind a joint team from Nanjing University and FiveAges.

The researchers used multiple instances of the lightweight Gemini 3 Flash model and reported better performance than the more powerful Gemini 3.1 Pro Preview benchmark. The results indicate that coordinated smaller models can outperform a single larger model on some physical navigation tasks, while destination verification can make embodied AI less prone to confidently stopping in the wrong place.