Swarms: Keep them simple, keep them social

Allowing neighbouring swarm robots to communicate and adapt their behaviour to a situation improves the swarm’s performance of an exploration task and outperforms random search methods.

White drone (also know as an UAV: unmanned aerial vehicle ) with four rotor blades flying in blue sky

Project team: Elliott Hogg, David Harvey, Sabine Hauert, and Arthur Richards

Keep it simple, swarms

Keeping technology simple is often a good maxim. This can be helpful with swarm robots, as simple robots are usually more robust than their complex equivalents.

The most basic exploration robots detect collisions with other robots and features, such as walls or boundaries, as they move through the environment.

Such robots can perform search tasks by acting independently. Better performance is obtained by allowing communication between robots to help them move away from each other. However, this leads to difficulties with certain environmental features, such as bottlenecks forming at doorways.

A behavioural approach is needed to maintain the elegance of simple technology while allowing robots to adapt their behaviour to different settings.

Happiness leads to positive adaptation

We tackled this need by introducing ‘happiness’ as an estimate of how well robots were exploring randomly generated simulated environments.

Each robot’s happiness increased with:

  1. how far the robot had travelled from its former locations;
  2. the density of nearby robots; fewer neighbours yields greater happiness.

A threshold value was applied to define a robot as either happy or unhappy at any point in time.

The key to robot happiness

  • Longer distances from past locations
  • Fewer neighbouring robots.

Robots would move in a particular direction until their unhappiness dropped significantly, either from encountering an obstacle or a group of other robots. The robot would then decide its new direction:

  • If other robots were nearby, they would communicate their level of happiness. The unhappy robot would then adopt the travel behaviour of a nearby happy neighbour;
  • If no other happy robots were nearby, the robot would move in a new, random direction.

This final step allowed one robot to use another’s happiness to find in which direction they should travel to continue exploring efficiently. This allowed robots to travel in the correct direction without complex mapping.

Exploration tasks were compared to the performance of robots using benchmark methods of either individual random walks or dispersion behaviour to move away from other robots.

Social robots work together

The swarms of social robots significantly outperformed the benchmark approaches.

The key to this improvement was that happy robots could guide unhappy neighbours towards regions that were more likely to offer greater opportunities for exploration. This sharing of happiness and adapted behaviour even saw robots form small teams that adopted the most appropriate behaviours to explore the environment quickly.

The simulations showed collaborating robots could adapt to complex and unknown environment geometries, and didn’t become deadlocked around features such as doorways.

Simple, happy robot swarms

Our simulations of self-supervising robot swarms show how effective local information sharing can be in improving exploration tasks. Replicating this behaviour in the real world would require robots with only basic levels of sensing and close-range communication, and without any centralised control or learning algorithms.

Overall, our work shows that distributed communication and simple decision-making rules allows robot swarms to perform exploration tasks efficiently and reliably.