R2BC: Multi-Agent Imitation Learning from Single-Agent Demonstrations

Researchers have introduced new methods for training multi-robot systems and improving multi-agent imitation learning, according to two recent studies published on arXiv^[1][2].

The first study introduces Round-Robin Behavior Cloning (R2BC), a method that enables a single human operator to train multi-robot systems through sequential, single-agent demonstrations. Imitation Learning (IL) is considered a natural way for humans to teach robots, particularly when high-quality demonstrations are easy to obtain^[1]. R2BC allows the human to teleoperate one agent at a time and incrementally teach multi-agent behavior to the entire system. The researchers showed that R2BC matches and surpasses the performance of an oracle behavior cloning approach in four multi-agent simulated tasks. Additionally, R2BC was deployed on two physical robot tasks trained using real human demonstrations. In a separate study, researchers presented a theoretical analysis of multi-agent imitation learning in linear Markov games. They proposed a computationally efficient interactive MAIL algorithm for linear Markov games, which has a sample complexity that depends only on the dimension of the feature map^[2]. The algorithm was shown to outperform Behavior Cloning (BC) on games like Tic-Tac-Toe and Connect4.