HUMAC system taught robots to anticipate teammates’ intentions
Researchers from Duke and Columbia University developed the HUMAC system, teaching robots team interaction through the Theory of Mind concept. The technology allows machines to anticipate partners’ intentions, forming the basis for coordinated cooperation.
Key efficiency indicators demonstrate a cool breakthrough. Without training, robots achieved success only in 36% of cases. After 40 minutes of training with a human curator, efficiency skyrocketed to 84% in simulation and 80% in real conditions.
The Human-guided Multi-Agent Collaboration methodology allows 1 person to teach a group of robots complex strategies – from encirclement to ambush. Brief hints are integrated into algorithms, forming a partner behavior model for each team member.
Practical tests were conducted in a hide-and-seek game. 3 seeker robots opposed 3 fast hiders in a confusing space with limited visibility. Results showed the system’s ability for intuitive action prediction without direct instructions.
Such robot capability for autonomous coordinated behavior opens possibilities for logistics, search-and-rescue operations and industry. However, the same principles are applicable for military purposes, creating ethical dilemmas of autonomous group decision-making.
