Communications Mean-shift exploration of shape assembly in robot swarms
Shape assembly of robot swarms has been extensively studied due to the fascinating collective behavior of biological systems6,7,8,9. In the literature, a class of strategies is based on goal allocation in centralized or decentralized ways10,11 and 12. After a swarm is assigned unique goals in the desired shape, it’s just a matter of planning collision-free routes for the robots so they can reach their goal positions10, or conducting distributed formation control using locally sensed data6,13,14. The computational complexity of a centralized goal assignment increases as the number and size of robots increase15,16. In addition, when robots do not function as expected, additional algorithms are needed for fault-tolerant detection, and goal reassignment to handle these situations17. Distributed goal assignment, on the other hand, can support large-scale flocks by breaking down centralized assignments into local ones11,12. This also makes it more robust to robot faults. Since distributed goal assignment is based on local information, conflicts between local assignments are inevitable. They must be resolved using sophisticated algorithms, such as local task switching11,12.
Another class of strategies for shape assembly that have also attracted extensive research attention are free of goal assignment18,19,20,21. The method described in ref. 18 can assemble complicated shapes by using thousands of homogeneous robotic arms. This method has an interesting feature: it doesn’t rely on external global position systems. It establishes instead a local position system based upon a small number pre-localized robots. Due to the local positioning system the proposed edge-following method of control requires that only robots at the edges of a swarm are allowed to move, while those within must remain stationary. The method described in ref. The 19th can produce swarm shapes from a reaction diffusion network, similar to the embryogenesis of nature. This method cannot generate shapes that are precisely specified by the user. The method described in ref. 21 can aggregate robots based on saliency. A digital light projector specifies the user-defined form. This method has the interesting feature that it doesn’t require edge detectors to be centralized. Edge detection is instead realized by combining the beliefs of robots with their neighbors. Since the robots are unable to self-localize relative to the desired form, they use random walks to find the edges. This would result in random trajectory. Artificial potential fields are another class of methods which do not require goal assignments22,23.24.25. This class of methods has the limitation that robots can easily become trapped in local minima. It is difficult to assemble complex nonconvex shapes.
We propose a strategy to shape-assemble robot swarms that is based on mean-shift exploration. When a robot finds itself surrounded by other robots, and in an unoccupied location, it will actively leave its current position by searching for the most unoccupied places nearby. This idea is not dependent on goal assignment. This is achieved by adapting the mean shift algorithm26,27.28. This is an optimization technique that is widely used in machine-learning for finding the maxima of density functions. A distributed negotiation mechanism allows robots to negotiate with their neighbors the final desired form in a distributed way. The negotiation mechanism allows the swarms to maintain a desired shape by relying on a few informed robots. The strategy allows robot swarms assemble complex nonconvex shapes with high adaptability, and this has been verified through numerical simulations and real-world tests with 50 ground robots. The strategy can be modified to produce interesting behaviors such as shape regeneration, collaborative cargo transportation, or complex environment exploration.
Source:
https://www.nature.com/articles/s41467-023-39251-5
