Onboard Evolution of Understandable Swarm Behaviors

Abstract

Designing the individual robot rules that give rise to desired emergent swarm behaviors is difficult. The common method of running evolutionary algorithms off‐line to automatically discover controllers in simulation suffers from two disadvantages: the generation of controllers is not situated in the swarm and so cannot be performed in the wild, and the evolved controllers are often opaque and hard to understand. A swarm of robots with considerable on‐board processing power is used to move the evolutionary process into the swarm, providing a potential route to continuously generating swarm behaviors adapted to the environments and tasks at hand. By making the evolved controllers human‐understandable using behavior trees, the controllers can be queried, explained, and even improved by a human user. A swarm system capable of evolving and executing fit controllers entirely onboard physical robots in less than 15 min is demonstrated. One of the evolved controllers is then analyzed to explain its functionality. With the insights gained, a significant performance improvement in the evolved controller is engineered.

Machine learning for active matter

“…Active-matter research should seek analogies with the problems studied in swarm robotics 21,77 , where genetic algorithms 21,22 (Box 1) and reinforcement learning 23 have been applied for some time. Recent trends in swarm-robotics research are deep reinforcement learning 78 , model-based behaviour trees evolved using genetic programming that result in humanreadable output 79 , and machine-learning methods based on generative adversarial networks 80 (Box 1 and Fig. 5b).…”

A Survey on Swarming With Micro Air Vehicles: Fundamental Challenges and Constraints

“…For these reasons, it is not surprising that, to the best of our knowledge and as confirmed by Chung et al (2018), most real-world implementations of MAV swarms to date have relied on primarily manually designed swarming algorithms. These advantages have also been acknowledged by the automatic design community, which has brought a general interest in using automatic approach to develop explicit controllers, such as state machines (Francesca et al, 2014(Francesca et al, , 2015 or behavior trees (Kuckling et al, 2018;Jones et al, 2019). In future work, the use of these methods could lead to a compromise between extracting an understandable controller and exploiting the power of automatic methods.…”

“…Moreover, thanks to the blind optimization, not only the controller can be evolved, but also other factors, such as the communication between robots (Ampatzis et al, 2008). Likewise, ER offers a generic approach to generate swarm controllers of different types, including, but not limited to: neural networks (Trianni et al, 2003;Silva et al, 2015), grammar rules (Ferrante et al, 2013), behavior trees (Scheper et al, 2016;Jones et al, 2018Jones et al, , 2019, and state machines (Francesca et al, 2014). Although neural network architectures can be very powerful, the advantage of the latter methods is that they can be better understood by a designer, which makes it easier to cross the "reality gap" between simulation and the real world when deploying the controllers on the real robots (Jones et al, 2019).…”

Distributed Situational Awareness in Robot Swarms

“…The resulting distributed situational awareness could result in richer information at the level of the swarm than with central situational awareness, with actions happening at the right location and time, leading to desired collective behaviors. [8] Distributed situational awareness fundamentally is more than the sum of its parts. Pooling all the information from the robots to a central control and then using this central control to direct the actions of the individual robots would invariably lead to a loss of information.…”