Category: Publication Research

Researchers from ISIR, Sorbonne University, and the Gulliver laboratory, ESPCI Paris – PSL, have designed robots that move in or against the direction of the forces to which they are subjected, in particular during collisions, based solely on their morphology. This work comes from the “Morphofunctional Swarm Robotics” project (ANR-18-CE33-0006) and is presented in an article published in the journal Science Robotics on February 22, 2023.

A swarm of robots can accomplish tasks requiring cooperation between several individuals. To date, swarms of robots are designed to operate exclusively in a diluted environment, i.e. avoiding collisions. However, this is not the case in living organisms, where great flexibility can be observed in both dilute and dense environments, be it for cells, colonies of bacteria or ants, shoals of fish, flocks of birds, or even humans. This suggests that simple primitive behaviors are sufficient to obtain collective behaviors, even in dense environments. Physical interactions can thus be used to the advantage of a swarm, even if it is composed of robots. The study and exploitation of these physical interactions between robots is the subject of an article at the interface of robotics, computer science and physics published in Science Robotics on February 22, 2023 by researchers from ESPCI Paris – PSL and Sorbonne University.

In this paper, the authors take advantage of a generic mechanical response: the tendency of a particle to reorient itself in response to an external force. They reveal the importance of this morphological response in different robotic tasks, both for a single robot and for a swarm of robots. The authors show that the way robots react to collisions depends on their body and allows them to align themselves with or against an external force such as another robot, an object or a wall. By understanding the effect of physical interactions between the robots and their environment, it becomes possible to define the collective behavior that will result from multiple collisions between robots in the swarm. Depending on the exoskeleton that dresses the robots, we can thus observe that one robot will align itself to an external force while another will oppose it. In case of collision, a robot can push or slide along an obstacle or another robot, and this only thanks to the passive dynamics induced by its mechanical design.

The authors of this paper have also shown that it is possible to learn to exploit interactions between robots to perform collective robotics tasks requiring cooperation, such as collective aggregation in a bright area of the environment. To this end, the authors describe a decentralized reinforcement learning algorithm inspired by social learning: sufficiently close robots exchange information about the most efficient behavioral strategies to accomplish the task assigned to the robot swarm. The best strategies are thus diffused from robot to robot, allowing the best use of the physical properties and computational capabilities of the robots in the swarm.

Caption: Cooperation between 64 Morphobots for a phototaxis task (search for light). Morphobots that are in the illuminated area form an aggregate that grows over time, exploiting the properties given by their morphologies. In case of a collision, each Morphobot naturally reorients itself to face the object at the origin of the force suffered. The formation of an aggregate in the luminous zone simply results from a slowing down of the speed of movement of the Morphobots, which takes place if the luminous intensity exceeds a threshold learned over time. The morphology does the rest. ©Matan_Yah_Ben_Zion

Authors (in order):