British artificial intelligence engineers have created a robot that mimics the evolutionary process in just the span of a few hours. They programmed a robot’s “brain” to automatically grow in size and complexity as its physical makeup is modified.
The legged robot (the image to the left shows a version with wheels), about the size of a paperback book, has a neural network (software) that grows by assigning new neuron clusters on top of existing structures as new limbs are attached. This approach is similar to the biological version, except brain complexity in animals grows steadily over millions of years.
The project's lead engineer, Christopher MacLeod, told New Scientist, "If we want to make really complex humanoid robots with ever more sensors and more complex behaviours, it is critical that they are able to grow in complexity over time-just like biological creatures did."
The team started with a simple neural network consisting of a module of four neurons with outputs connected to the robot's leg actuators. Then, they wrote an "evolution" algorithm that trained the network to produce desired actions, such as remaining balanced or travel a certain distance. According to the engineers:
The robot (and the environment it interacts with) is allowed to become slightly more complex and the process of adding further network modules is repeated. This process continues until the robot can fulfil all its desired tasks. The previously added functionality stays (it does not evolve further, only the newly added modules evolve) and the newer parts build up like the layers of an onion.
As the quadruped robot learned to walk, Macleod said, "It fell over mostly in a puppyish kind of way. But then it started moving forward and not falling straight away-and it got better and better until it could eventually hop along the bench like a mudskipper."
The team tested ways to "re-evolve" the entire brain but that crashed the software. Work is underway to develop the network to handle more versatile limbs, to integrate a sensory network, and create an algorithm that controls the overall development of the system.
Previous research associated with artificial neural networks stopped short of configuring large networks which fuse data from various sensors in a complex and open-ended manner. You can download a paper (PDF) from The Robert Gordon University website for details of the research, including charts depicting the evolution of the robot's gait.
Macleod thinks that the research will help build more advanced robots, and can potentially find applications in humanoid robots and prosthetic limbs.
Experimental evolution with robots is not limited to locomotion. Some may recall that last year, scientists at the Swiss Federal Institute of Technology demonstrated how robots can evolve to communicate with each other, to help, and even to deceive each other.