Robot swarms 'evolve' effective communication

 作者:嵇丽     |      日期:2019-03-01 07:18:10
By Tom Simonite (Image: Current Biology) Robots that artificially evolve ways to communicate with one another have been demonstrated by Swiss researchers. The experiments suggest that simulated evolution could be a useful tool for those designing of swarms of robots. Roboticists Dario Floreano, Sara Mitri, and Stéphane Magnenat at the Swiss Federal Institute of Technology in Lausanne collaborated with biologist Laurent Keller from the University of Lausanne. They first evolved colonies of robots in software then tested different strategies on real bots, called s-bots. Both simulated and real robots were set loose in an arena containing two types of objects – one classified as “food” and another designated “poison” – both lit up red. Each bot had a built-in attraction to food and aversion to poison. They also have a randomly-generated set of parameters, dubbed “genomes” that define the way they move, process sensory information, and how they flash their own blue lights. “They start with completely random behaviour,” Keller explains. “All they can do is discriminate food from poison.” The robots can see both food or poison from a distance of several metres but can only tell them apart when almost touching. The simulated bots were allowed to explore their surroundings. The genomes of the bots that found food and avoided poison most efficiently were recombined, mimicking biological natural selection. Their “genomes” were combined and randomised in a way designed to mimick mating and mutation and used this to create the next generation robot. In each experiment, five hundred generations were evolved this way, under different selective pressures. “Under some conditions, sophisticated communication evolved,” says Keller. “We saw colonies that used their lights to signal when they found food and others that used signals to communicate they had found poison.” Cooperative communication evolved when selective success was judged at the group level – when many robots displayed efficient behaviour – or when the genomes of the robots were most similar – like biological relatives. Real s-bots with genomes evolved through simulations displayed the same behaviour. This video shows the robots in action. The same s-bots have previously been used to test different cooperation strategies (see Robot swarm works together to shift heavy objects). Genes that promote sharing spread when an individual’s survival is tied to that of the whole group – like in bees or ants. They also do well in populations of close relatives, since helping kin helps the same genes reach the next generation. In contrast, in simulations with bots not closely related and when group performance was less important, some robots evolved misleading behaviour. Some even produced signals that lured other bots away far from food. Further experiments involving real robots will be used to investigate ways that evolution could be used as a practical design tool. Keller also plans to test what happens when evolved and un-evolved bots mix. “These are significant findings for the biologically-inspired robotics community,” says Noel Sharkey, who is also evolving robot behaviour at Sheffield University, UK. Learning how to evolve robots instead of writing their behaviour from scratch could ultimately lead to more sophisticated behaviour, he says. However, artificial evolution has yet to mimic all the powers shown in biology. Sharkey points out that his robots already start with a means of communication. “In real biology no such communication channel exists initially and has to be created rather than discovered,