Portuguese scientists have built robots that aren't programmed to perform tasks in the normal sense. They use algorithms that allow them to evolve their own behaviours by learning online.
The robots were even able to overcome faults in their motors, and adapt to changes in the tasks they were set.
The team successfully evolved neural network-based controllers in real robotic hardware to solve two single-robot tasks and one collective robotics task.
Controllers are evolved either from random solutions or from solutions pre-evolved in simulation. In all cases, solutions are found in less than an hour.
This is the first time the adaptive capabilities of online evolution have been shown in real robotic hardware. It is the first time robots have been shown to overcome faults injected in the motors of multiple units simultaneously, and it is the first time robots have been able to change their behaviour when the task did.
They even have a plan for further evolution for more complex tasks.
"For general-purpose robots capable of learning a variety of different skills for different tasks to become feasible, it is required that they can learn new skills while retaining old skills, " the researchers say. "However, such an objective remains elusive because robots and artificial agents tend to forget previously learned information in a catastrophic manner when multiple skills are required."
"One solution to allow robots to accumulate skills throughout their lifetime is to rely on modular control systems in which different skills are represented as building blocks of the system, and then combined in order to solve more complex tasks."
They have taken the first steps towards this goal - developing an approach in which building blocks are evolved or preprogrammed "and then seamlessly specified in the neural structure. The building blocks are optimised together with the neural networks' weights and topology in a unified manner."
The team plans to explore the usage of this and similar techniques accumulate learned behaviours and techniques for problem solving.
Then: global domination. Right?! [Source]