A group of robots can now detect abnormal behavior within its ranks and coordinate its response without a central computer directing them.
A robotic swarm relies on many small machines that each make decisions based on local information. No robot necessarily has a complete view of the group. This organization avoids depending on a single leader, but it creates a difficulty: how can the group know what it is really doing when each member sees only part of it?
A large swarm of Kilobot robots, a platform designed to study the collective behaviors of many small machines.
Credit: Asuscreative / Wikimedia Commons — CC BY-SA 4.0
Researchers at the University of Utah have developed a method called SyncSBC to address this question. Each robot observes its neighbors and uses this limited data to estimate the collective behavior. The machines then exchange their estimates to reach a common decision, without transmitting all their data to a central control station.
The system combines machine learning with exchanges between robots. The former is used to recognize different swarm behaviors from the available observations. The exchanges then allow the group members to quickly align their decisions. As a result, each machine can form an estimate of the overall behavior without directly observing all the other robots.
This capability becomes particularly useful when a group member no longer behaves as expected. A failure, an incorrect command, or an unexpected change can disrupt the collective movement. During tests on real robots, SyncSBC enabled the group to identify behavioral anomalies using only its distributed observations.
The researchers also tested a second possibility: collectively changing the swarm’s behavior. Once a situation is recognized, the robots can synchronize their decision and autonomously switch to a different organization. No central server needs to decide on the change for the entire group.
This lack of a command center offers an advantage. In an environment where communications are limited, distant, or unstable, a permanent connection to a single computer can become a weak point. An autonomous swarm can continue making certain decisions even when each robot has only information from its immediate surroundings.
The results were obtained using real robots, but the work remains experimental research. The authors report high classification accuracy and a short synchronization delay in their tests. The code, videos, and additional experiments have been made available to allow the method to be studied in other swarm configurations.