ULSAN — Researchers have developed artificial intelligence technology that lets different robots, such as drones and walking robots, share visual experience gathered in the field so they can accurately gauge distances even in unfamiliar surroundings. Because raw camera footage never leaves the robot, the method sharply improves autonomous navigation accuracy while easing privacy risks and communication costs.
Ulsan National Institute of Science and Technology said on the 3rd that a team led by Professor Joo Kyung-don of its Graduate School of Artificial Intelligence, working with a team led by Professor Yoon Sung-hwan, developed a federated learning technique called FeDepth that precisely trains AI models for monocular depth estimation in robots.
Monocular depth estimation calculates the distance between objects and spaces using video from a single camera, without expensive and complex sensors. Robots need to learn from vast amounts of data to safely recognize routes in unknown environments, but conventional approaches that gather high-resolution video on a central server carry heavy communication costs and the risk of privacy breaches. Federated learning, which exchanges only training results rather than raw data, emerged as an alternative. Yet performance degraded when data from robots with entirely different platforms, fields of view and surroundings — such as an airborne drone and an indoor serving robot — was simply averaged together.
To overcome that data heterogeneity, the team adopted soft clustering. Summary information on each robot's visual characteristics is compared so that robots that have experienced similar environments train together, but no robot is locked into a single group; instead, each takes part in multiple related groups. A drone that has flown over a forest, for example, forms one group with a quadruped walking robot that has moved through the same forest while also forming a separate group with a drone that has flown over a city, sharing training across both. That blocks distortion from conflicting data while capturing all shared spatial characteristics in the model.
In verification tests, FeDepth reduced the relative error in distance estimation by about 17% to 32% compared with standard federated learning under conditions where robot platforms and driving environments were mixed. In tests against the conventional single-group assignment method, the relative error fell about 6%, to 0.249 from 0.264. The technique showed even performance gains from short-range indoor settings within 10 meters to long-range outdoor settings of up to 80 meters, and delivered the same improvement across all three monocular depth estimation AI models of differing architecture.
"FeDepth achieves efficient joint training that reflects differences in robot models and operating environments without exchanging raw video," the team said. "It will raise spatial recognition capabilities by one level across a range of unmanned mobility fields, including self-driving cars, logistics and delivery, service robots, disaster rescue and forest monitoring."
Researchers Lee Kang-hyun and Lee In-ha participated as first authors. The findings were accepted by the European Conference on Computer Vision (ECCV 2026), the most authoritative international conference in computer vision, and will be formally presented in Malmo, Sweden, from Sept. 8 to 12. The research was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) under the Ministry of Science and ICT and the National Research Foundation of Korea (NRF).
