Korean researchers have developed technology that allows multiple cooperating artificial intelligence (AI) systems—such as autonomous drones or smart factory robots—to complete missions on their own by redistributing roles, even when some devices fail or lose communication.
A team led by Professor Han Seung-yul at the Artificial Intelligence Graduate School of the Ulsan National Institute of Science and Technology (UNIST) said Tuesday it has developed a multi-agent reinforcement learning technology called Interaction-Breaking Adversarial Learning (IBAL), which trains AI by deliberately severing the cooperation links between AI agents.
The IBAL system devised by the team can be compared to a soccer match. When a player is sent off during a game, rather than sticking to the existing tactics, the remaining players are trained to immediately fill the empty space and redistribute their offensive and defensive roles.
Conventional multi-agent reinforcement learning mainly used methods that deliberately injected noise into sensor information or induced individual AI to take actions unfavorable to the mission. By contrast, IBAL divides an AI group into two groups, analyzes the "mutual information" most critical to their cooperation, and intensively disrupts it.
During the learning process, the method either obscures the "information" essential for AI to understand one another, or induces "behavior" that breaks cooperation. It is also designed so that AI broadly experiences various forms of cooperation breakdown and builds response capabilities, by randomly changing the group composition at each learning step and automatically adjusting the intensity of the attacks.
The team verified the technology's performance in an experimental environment (SMAC) based on the famous strategy game "StarCraft II."
When assuming a sudden situation in which some friendly units abruptly stop functioning, existing AI models saw their entire cooperation system collapse like dominoes, with win rates plunging to as low as 13.3 percent. By contrast, AI trained with IBAL immediately rebuilt its formation—pulling units with depleted health to the rear and putting healthy units at the front—recording an overwhelming win rate of 87.0 percent.
"IBAL does not stop at interfering with the judgment of individual AI, but shakes the cooperative relationships among AI themselves," said lead author Lee Sun-woo, a researcher. "Through this, we can train the remaining AI to find new ways of cooperating and continue the mission, even when some AI fail or lose communication."
"In systems where multiple AI move together, such as autonomous drones, swarm robots, and smart factories, fatal problems arise when some equipment fails or communication is cut off," Professor Han Seung-yul stressed. "This technology will become a core foundational technology for enhancing the overall safety and reliability of systems in which multiple AI operate together."
Meanwhile, the research results were accepted to the International Conference on Machine Learning (ICML) 2026, one of the world's three leading AI conferences. This year's ICML will be held at COEX in Seoul from July 6 to 11. The research was conducted with support from the Ministry of Science and ICT, the Institute of Information & Communications Technology Planning & Evaluation (IITP), and the National Research Foundation of Korea (NRF).
