Collective movement is a widely observed phenomenon in nature. Be it schools of fish, flocks of birds or swarms of insects, animals can implement well-coordinated collective movement relying only their perception, i.e. the locally available information surrounding them. Current flocking models often rely on, instead of local, global information such as the exact position of conspecifics. How does collective motion arise only from local perception? Is it possible to create a purely vision-based controller for robots that implements similarly coordinated movement?

To study these questions we implemented a novel vision-based flocking model introduced by Bastien and Romanczuk in both simulated agents and in 10 small terrestrial robots. We studied the following:

• How does a limited FOV influence the original vision-based flocking model?
• How does the model perform in a constrained environment (walls)?
• How does the model perform on real robots with a fully decentralized design?

We found in simulations that limiting the visual perception of individuals fundamentally changes the resulting movement patterns, and less visual information sometimes even favors coordinated movement. Combining our findings with limited spaces, on the other hand, always made it more challenging for groups to realize coordinated migrating movement without collisions as in these cases agents are continuously perturbed by limiting walls.

We also found that the smaller the agents' FOV is the less cohesive the group remains as agents can break away from the group if no other agent is visible.

Using our results from simulations involving agent-based models, we developed a vision-based controller for small robots. This controller enables the robots to achieve coordinated movement without the need for any centralized control system. The robots can move in unison by solely relying on their camera feeds to process all necessary information on-board, including the detection of each other through CNN-based object detection methods. Decentralized control could offer significant advantages over centralized systems in the future. Robots that rely solely on local information are more resilient to disruptions in global information systems, such as GPS jamming or outages. Moreover, the agents do not require direct communication with each other, facilitating coordinated movements among various types of robots without the need for a compatible communication protocol.

Manuscript is coming soon...

• GitHub Find the Code on GitHub

Videos