7 August 2015 – Swarm robotics deals with the control of large groups of relatively simple robots so that they perform tasks that go beyond their individual capabilities. Many biological systems execute tasks by dividing them into finer sub-tasks first, may be you have seen Ants working together in a large scale to achieve a common big goal. It looks very interesting but more than that it is very challenging job. By seeing them first question comes in our mind that how they are dividing work among themselves and how they are coordinating to each other? In other words, how do they self-organize their tasks?
To answer this question, Eliseo Ferrante and team from the University of Leuven in Belgium have developed a novel method that allows robots to efficiently self-organize their tasks in robot swarm.
A limitation of traditional analytical modeling approaches to division of labor, however, is that they can only consider a finite and pre-specified number of behavioral strategies. In recent years, artificial evolution of teams of embodied agents has been used to enable the study of social traits in more detail, taking into account more realistic physical constraints and a much larger set of allowable behaviors and strategies. In evolutionary swarm robotics, no study has succeeded in evolving complex, self-organized division of labor entirely from the beginning. This may be due to the fact that most evolutionary robotics studies have made use of neural network-based approaches, which have been shown to scale badly to more complex problems.
In this study, Researcher first time indicated that a self-organized division of labor mechanism could be evolved entirely from the beginning. They used simulated teams of robots and artificially evolved them to achieve maximum performance in a foraging task.
The main aim of this study was to test if other nature-inspired evolutionary methods than traditionally used in evolutionary swarm robotics would be able to achieve complex task specialization in social groups. This nature-inspired evolutionary method allows a set of low-level behavioral primitives to be recombined and evolved into complex, adaptive behavioral strategies through the use of a generative encoding scheme that is coupled with an evolutionary process of mutation, crossover and selection.
They demonstrated result with “task partitioning” example. The setup was like leafcutter ants, whereby some ants (“droppers”) cut and drop leaf fragments into a temporary leaf storage cache and others (“collectors”) specialize in collecting and retrieving the fragments back to the nest. In analogous robotics setup, they used a team of robots simulated in-silico using an embodied swarm robotics simulator and required the robots to collect items and bring them back to the nest in either a flat or sloped environment.
The results of these experiments show for the first time that complex, self-organized task specialization and task allocation could be evolved in teams of robots. Nevertheless, a fitness landscape analysis also demonstrates that task specialization was much easier to evolve when pre-evolved behavioral building blocks were present.
(From initially random behavior, the robots first evolve generalist foraging after 500 generations. Subsequently, after 500 more generations, the robots evolve task partitioning, which gets further perfected over the following 1000 generations. We conclude by showing how the controller evolved in a 4 robot team scaled up when tested in a swarm of 20 robots.)
- Explore further: Buzz: A New Programming Language for Robot Swarm
- Reference: “Evolution of Self-Organized Task Specialization in Robot Swarms”, Eliseo Ferrante, Ali Emre Turgut, Edgar Duéñez-Guzmán, Marco Dorigo, Tom Wenseleers Published: August 6, 2015 DOI: 10.1371/journal.pcbi.1004273
- Image: 1) Eliseo Ferrante/University of KU Leuven 2) From Paper