Survival of the fittest digital brain

by Lakshmi Sandhana  A holistic, evolutionary approach means that robots could learn to design themselves WHAT if a robot – brain, body and all – could be born and then develop in a similar way to a human baby? Instead of…

by Lakshmi Sandhana 

A holistic, evolutionary approach means that robots could learn to design themselves

WHAT if a robot – brain, body and all – could be born and then develop in a similar way to a human baby?

Instead of a mother, the robot would come out of a printer, in its entirety. “The end game is to evolve robots in simulation, hit print, and watch them walk out of a 3D printer,” says Jeffrey Clune, who heads the HyperNEAT project at Cornell University’s Creative Machines Lab in Ithaca, New York.

Before he can do that, though, Clune needs the right design. His team had already evolved digital brains using neural networks that mimic biological evolutionary processes, and the researchers are now connecting these brains to a body, to find ones that can make the body walk right away.

The neural networks, essentially a series of algorithms, enable the brain to learn how to control physical robotic bodies, either simulated or physical. The brains receive sensory inputs from the body telling them what to learn to control – whether it has two legs or four, for example – then evolve the neural patterns needed to control it.

Each brain was given control of a physical body for a period of time. Some were abject failures, and could only muster enough control to flail about, fall over or twitch.

The best-performing brains were allowed to reproduce to create the next generation and the entire process was then repeated until the team obtained a brain that could control the robot and walk around the lab. So far, Clune’s team has evolved a brain that is able to make a four-legged robot walk within a few hours of the brain being plugged into the body. The results were presented last month at the European Conference on Artificial Life in Paris, France.

“From an observer’s perspective, it looks like a robot that ‘wakes up’, tries out a new gait, and then ‘thinks about it’ for a few seconds, before waking up again and trying a new gait,” says Clune. “Over time you see that the robot learns how to walk better and better.”

As the robot brains can adapt according to the information given them from the body, they can learn continually and transfer acquired skills from one task to another. For example, a robotic brain evolved to control a four-legged robot would still function if hooked up to a six-legged robot. A small amount of damage shouldn’t cripple these robots as they would be able to adapt.

Clune’s team is now evolving simulated bodies and brains with theirEndlessForms website, also developed at Cornell. This uses evolutionary algorithms to gradually modify designs before bringing them into the real world with 3D printing. Clune hopes to use EndlessForms to design soft-bodied robots using printable materials that act as muscles, bones, batteries, wires and even computers.

The lab has 3D-printed many of these components already, including wires and artificial muscles, which move when a current is passed through them. However, they have yet to find a way to print structural material with different levels of stiffness – harder materials for bone, for instance – as well as some of the softer, more flexible tissues. “Eventually, the entire thing will be printed, brains and all,” says Clune.

The team uses neural networks known as compositional pattern-producing networks (CPPNs) to mimic how natural organisms develop. This produces designs that share important properties with natural organisms, such as symmetry and the repetition of modules.

Josh Bongard, who works on robot evolution at the University of Vermont in Burlington, says Clune’s approach is exciting because it explores how a robot’s body affects its behaviour and offers control over the evolution of every aspect of a robot – brain, body and behaviour.

“If CPPNs are used to evolve robot bodies along with brains, he may be able to evolve robots with complex bodies as well as complex brains,” Bongard says.

Using evolutionary methods to build whole robotic entities opens up an even more intriguing possibility – that robots could evolve new and entirely different structures. “It may well be that the neural networks and bodies that we create for the system are not those that the system would develop for itself,” says James Giordano, who directs the Center for Neurotechnology Studies in Arlington, Virginia.

Better robotic bodies that can handle the brain’s demands are certainly needed: the test robot broke down because it was unable to cope with the running motions that the brain had evolved after a number of generations.

Eventually, the brain and body should work in concert, with the brain’s evolution dependent upon what tasks the robot carries out during its lifetime, says Jean-Baptiste Mouret of the Intelligent Systems and Robotics Institute in Paris, France.

“The brain will depend on the body and on the ‘life’ of the robot, in the same way as birds’ brains are different from rats’ brains,” says Mouret. Breeding robots that can do more than walk is the next big step. The software-based brains that are best at performing a desired behaviour in simulation, such as climbing a wall or getting close to a person, will be allowed to reproduce until the final generation has hard-wired instincts to perform the task.

Understanding intelligence

The “brain plus body” approach to evolving robots (see main story) may also help researchers in other fields, allowing them to study the conditions under which intelligence evolves fastest. For instance, one theory says that intelligence evolved to keep track of life in complex social societies. This can be tested by studying robot evolution when they are alone and in groups. “These are dream experiments for evolutionary biologists and they become possible using evolution in computers, but only once we are able to evolve complex-enough brains,” says creator Jeffrey Clune. “That’s one reason our breakthrough is so important: it opens doors to many new types of science.”