Researchers led by Ziqiang Liu have presented EgoSwarm, a framework that teaches groups of robots to cooperate without central control by learning from groups of people, according to a paper on arXiv. Robot learning usually relies on single-robot demonstrations, and collecting demonstrations with many robots at once is hard, the authors write.
They recorded more than 50 hours of people working together with cameras worn at eye level, without any robots, and recovered who was interacting with whom from the separate recordings. A model called SwarmVLA then turns this into policies that act on each robot's own view, its recent history and short messages from neighbours.
The policies succeeded in 72.2% of attempts on average across three categories of multi-robot collaboration tasks.



