Evolutionary Robots

I don’t know how relevant this work will prove to be in my own dissertation, but it’s cool enough to share. My friend Kynthia passed along a link to a paper on some successful experiments involving evolution in robot behavior.

The purpose of this work, inspired by Darwinian selection theories, is to illustrate how natural selection can lead to complex properties—in particular, adaptive behavior:

Just a few hundred generations of selection are sufficient to allow robots to evolve collision-free movement, homing, sophisticated predator versus prey strategies, coadaptation of brains and bodies, cooperation, and even altruism. In all cases this occurred via selection in robots controlled by a simple neural network, which mutated randomly.

One key insight from this work is that it highlights the perhaps flawed strategies in past attempts to make robots smarter. The authors point out that other techniques rely on vast prior knowledge of a particular environment in order to be successful. Failures occur when the robot shifts environments to find out a dependent resource is no longer there, or that it is incapable of detecting the information in the new setting that would be most helpful.

In a previous paper (Floreano, D., Husbands, P., and Nolfi, S. (2008). Evolutionary robotics. In: Siciliano, B., Khatib, O., editors. Springer handbook of robotics. Berlin: Springer Verlag. pp. 1423–1451.), Dario Floreano listed some of the ways engineers have attempted to address the problem of real-world failures with their robots:

• Measure the fitness of evolving individuals in several environments, varying dimensions in each
• Incorporate noise into the simulation models that moves the virtual environments away from the ideal
• Coevolve the robot and the key parameters of the simulation model, periodically testing with real robots to improve fitness estimations
• Add ontogenetic plasticity to the evolving individuals, so they can adapt to environmental changes over the course of their lifetime

I find this intriguing from a social robotics scale because of the dynamics of social interaction is often unpredictable. There may be some lessons to learn from this approach, even if I don’t end up building anything.

Reference:

• Floreano, D., and Keller L. (2010). Evolution of adaptive behaviour in robots by means of Darwinian selection. PLoS Biol, 8(1)