Reverse-engineering of Artificially Evolved Controllers for Swarms of Robots

It is generally challenging to design decentralized controllers for swarms of robots because there is often no obvious relation between the individual robot behaviors and the final behavior of the swarm. As a solution, we use artificial evolution to automatically discover neural controllers for swarming robots. Artificial evolution has the potential to find simple and efficient strategies which might otherwise have been overlooked by a human designer. However, evolved controllers are often unadapted when used in scenarios that differ even slightly from those encountered during the evolutionary process. By reverse-engineering evolved controllers we aim towards hand-designed controllers which capture the simplicity and efficiency of evolved neural controllers while being easy to optimize for a variety of scenarios.

Communication-based Swarming for Flying Robots

Dario Floreano, Sabine Hauert, Severin Leven, Jean-Christophe Zufferey

Swarms of flying robots can be used in disaster areas to autonomously create communication networks for rescuers and victims. Flying robots have the advantage of rapidly overcoming difficult terrain and providing unobstructed wireless communication. To allow for a swarm composed of cheap, transportable and robust robots, we avoid using positioning sensors which typically depend on the environment (GPS, cameras) or are expensive and heavy (lasers, radars). Instead, robot behaviors depend on local communication with robots within transmission range. There currently exists no methodology to design robot controllers resulting in the emergence of desired swarm behaviors. Here, we propose two bio-inspired techniques to overcome this problem. In the first approach, we use artificial evolution as a mean to automatically design simple, efficient and unthought-of controllers for robots. We then reverse-engineer these controllers and reuse the discovered principles in a wide variety of scenarios. In the second approach, we look at the creation, maintenance and evaporation of army-ant pheromone trails during foraging and apply the same principles to the design of robot controllers for the deployment, maintenance and retraction of communication networks.

2010

Evolution of cooperation in artificial ants

Markus Waibel

The evolution of cooperation is a fundamental and enduring puzzle in biology and the social sciences. Hundreds of theoretical models have been proposed, but empirical research has been hindered by the generation time of social organisms and by the difficulties of quantifying costs and benefits of cooperation. The significant increase in computational power in the last decade has made artificial evolution of simple social robots a promising alternative. This thesis is concerned with the artificial evolution of groups of cooperating robots. It argues that artificial evolution of robotic agents is a powerful tool to address open questions in evolutionary biology, and shows how insights gained from the study of artificial and biological multi-agent systems can be mutually beneficial for both biology and robotics. The work presented in this thesis contributes to biology by showing how artificial evolution can be used to quantify key factors in the evolution of cooperation in biological systems and by providing an empirical test of a central part of biological theory. In addition, it reveals the importance of the genetic architecture for the evolution of efficient cooperation in groups of organisms. The work also contributes to robotics by identifying three different classes of multi-robot tasks depending on the amount of cooperation required between team members and by suggesting guidelines for the evolution of efficient robot teams. Furthermore it shows how simulations can be used to successfully evolve controllers for physical robot teams.

EPFL2007

Evolutionary Synthesis of Communication-Based Aerial Swarms

Sabine Hauert

Aerial swarms have the potential to search for forest fires, chemical plumes or victims and serve as communication and sensor networks in the sky. Flying robots are interesting for such applications because they are fast, can easily overcome difficult terrain and provide line-of-sight communication or aerial perspectives. However, swarms of flying robots have so far only been demonstrated in simulation or in few examples in reality. Current simulators typically rely on unrealistic assumptions concerning robot sensing capabilities and motion. To bridge this reality gap we propose to address two key challenges. The first challenge consists in discovering swarm controllers that do not use position information. Swarm controllers in the literature rely on global or relative position information which can be obtained using GPS, cameras or range and bearing sensors. However, position requirements typically entail robots that are complex, heavy and expensive or that rely on specific environmental conditions to function. Lifting the need for position could instead lead to swarm controllers that can be deployed in a variety of environments and using very simple robots. The second challenge addressed consists in developing swarm controllers that can accommodate motion constraints of flying robots. In particular, we consider fixed-wing robots that can not stop or turn on the spot like ground robots or rotor crafts. Instead they must fly at relatively high speeds to avoid stalling. Therefore, making the robots advance slowly in average can only be done by having them turn. To address these challenges, we consider robot controllers that continuously react to wireless communication with neighboring robots or people by changing their turn rate (communication-based behaviors). Using communication as a sensor for flying robots is interesting because most robots are equipped with off-the-shelf radio modules that are low-cost, light-weight and relatively long-range. The design of robot controllers that can lead to desired swarm behaviors is done following a systematic approach. First, artificial evolution is used to automatically discover simple and unthought-of controllers for swarms of flying robots. Evolved controllers are then reverse engineered to obtain basic behaviors that can be modeled and used in a variety of swarm applications. In order to discover a range of basic behaviors, we consider a challenging swarm application that requires robots to direct themselves in their environment, move in groups, cover an area, and maintain a communication relay. Discovered behaviors are then extended to scenarios with wind. Overall, this thesis presents the evolutionary synthesis of communication-based behaviors for swarms of fixed-wing flying robots.

EPFL2010

Evolutionary Synthesis of Communication-Based Aerial Swarms

Sabine Hauert

EPFL2010