On-the-Fly Evolution for

Keywords: evolutionary , evolu- On-line, on-board adaptation tionary algorithms, modular robotics On-line is the keyword here: the EA monitors Imagine a collection of small, relatively sim- and refines the controllers as the robots ple, autonomous robots that collectively have go about their regular tasks. This is a major to perform various complex tasks. To achieve departure from ‘traditional’ evolutionary ro- their goals, the robots can move about indi- botics, where the controllers are developed vidually, but more importantly, they can off-line, and remain fixed once the robots are physically attach to each other to form and deployed actually to perform their tasks. manipulate multi-robot organisms for tasks An on-line, on-board EA has to address a that an unconnected group of individual ro- number of uncommon impositions such as bots cannot cope with. Think, for instance, of very limited resources, in terms of both proc- scaling a wall or holding a relatively large ob- essing power and memory space, a focus on ject. average rather than best performance and the One of the advantages of a of simpler need for on-line parameter control. robots is the increased robustness compared We have developed two algorithms specifically to complex monolithic systems: if a single to address these issues; an encapsulated (i.e., robot fails, the swarm can pretty much carry running within a single robot) and a distrib- on regardless. Also, the robots can reconfig- uted (over multiple robots) on-line EA. Both ure the organism to suit particular tasks and have been validated in a number of experi- circumstances, something that large and com- ments, with further experiments to combine them underway. Modular Organism Control Once joined into a larger ‘organism,’ the module controllers have to collaborate very precisely to manipulate the organism. We have developed generative encodings for this task: rather than evolving the controllers directly, the EA develops programs that generate con- trollers for individual robots joined in a larger organism. To allow for specialisation without sacrificing commonality, the same program generates controllers for all modules, but differentiating according to their position in the organism. plex individual robots would find impossible. Thus, a module in, say, an organism’s ‘spine’ is The SYMBRION/REPLICATOR project focusses similar to but different from one in the ‘leg.’ on developing techniques to allow the robots Experiments have shown this approach to to learn: to adapt their controllers to various yield effective modular control for organism tasks and circumstances. It forms one of the locomotion. largest projects in to- day, comprising a collaboration between Further information European universities from Germany, the For further information on the SYMBRION/ United Kingdom, France, Belgium, Austria REPLICATOR project as well as an overview of and the Netherlands. Within this project, the resulting publications, please refer to VU University Amsterdam’s Computational http://www.symbrion.eu or contact the Intelligence group researches and develops authors. evolutionary algorithms (EAs) for on-line, on- board adaptive robot control. A.E. Eiben and Evert Haasdijk Faculty of Sciences - VU University Amsterdam [email protected], [email protected] De Boelelaan 1081a +31 (0)20 5987758 1081 HV Amsterdam, The Netherlands