When people explore a new room, they roughly picture its layout. Robots usually work differently: they need a ready-made map or data from cameras and sensors. But what if the robot is tiny and has almost no sensors or memory? An international team of researchers from the Institute of Intelligent Systems and Robotics at Sorbonne University (ISIR), the University of Tokyo and other organizations proposed an answer. Their work was published in the journal Science Robotics.
The scientists asked whether a group of simple robots can recognize the shape of a space that none of them sees in full. The paper's first author, Leo Cazenille, compared this to the parable of the blind men and the elephant: each touches only part of the animal and cannot guess its overall appearance. Likewise, a robot perceives only its nearest neighbors, without GPS, a camera or enough computing resources to build a map.
The researchers were inspired by mathematician Mark Kac's question: "Can one hear the shape of a drum?" Sound depends on the shape of the membrane. The scientists suggested that a group of robots could likewise "listen" not to sounds but to the numerical information its members exchange. The new approach was called spectral swarm robotics. The word "spectral" here refers to mathematical analysis: it concerns a numerical "fingerprint" of the arena's shape.
Each robot stores a number representing a virtual temperature and exchanges it with neighbors, updating the value as if heat were spreading within the group. The rate at which the "temperatures" equalize depends on the network of connections between the robots and reflects the geometry of the space. By observing how its number changes, a robot computes a numerical signature of the geometry and compares it with known samples. The method requires only short messages and simple calculations - no cameras, central computer or information about its own position.
First the authors tested the approach mathematically and ran simulations with swarms of up to several hundred robots in seven arena shapes. Then followed experiments with real swarms of two types - Kilobots and Pogobots. In the experiments, both types successfully distinguished a circular arena from an annular one. Kilobots correctly identified the shape in 22 out of 30 cases. And a group of just 25 Pogobots handled the task in all 30 trials. The researchers also identified a relationship between swarm size, communication range and arena area that helps select optimal conditions.
According to Cazenille, the work shows that the pattern of connections between robots can itself serve as a tool for perception and computation. Although the algorithm targets arena-shape recognition, the approach may prove general enough for more complex tasks that require locally estimating global properties.
It is important to note: the study tested specifically the robots' ability to distinguish shape. How much this will improve swarm navigation remains to be assessed. In the future, the scientists plan trials in more diverse environments, want to teach robots to estimate several spectral values and to keep moving while the communication network changes. Also planned are tasks beyond shape recognition, such as identifying boundaries or cavities, and, in perspective, linking collective perception with action so that a swarm can change its formation in response to the environment it perceives.