Evolutionary robotics is a sub-field of artificial life in which evolutionary algorithms are used to evolve robot morphology (i.e. shape) and/or behavior . Sometimes these robots exist only in simulation, while other times evolved morphologies and behaviors are translated out of the computer and into physical robots. Getting robots evolved in a computer to work in the real world is notoriously challenging due to what is known as the “reality gap” (i.e. the many unexpected differences between a simulated environment and reality), but there has increasingly been progress towards overcoming this problem.
For a video introduction to evolutionary robotics, see the following talk by Sam Kriegman at the 2021 ISAL Summer school:
The first well-known instance of evolutionary robotics was the work of Karl Sims, who evolved life-like moving creatures in simulation out of rigid block-like components :
A few years later, Hod Lipson and Jordan Pollack used the same principles to evolve robots that could be physically constructed in real life :
Since then, the field has flourished and continued in a variety of directions.
One central challenge in evolutionary robotics is how to evolve effective locomotion behaviors (or gaits) for a given robot morphology. This problem is often tackled from an evolutionary computation perspective. Locomotion involves a variety of sub-problems:
Jared Moore and Adam Stanton have carried out a detailed set of investigations on a simulated wall-crossing task, for which they have found lexicase selection to be a particularly effective strategy .
Another common task in robot locomotion is evolving robots that are capable of adapting their gait when they are damaged. One recent prominent solution to this problem used the MAP-Elites algorithm to evolve a suite of gaits for a hexapod robot that used different legs to different extents .
Complex robot behaviors
There has also been a substantial amount of work on evolving more complex high-level of behavioral controllers for robots. One particularly notable example of this was the Twitch Plays Robotics project, in which a crowd of human observers functioned as the fitness function for evaluating simulated robots’ performances on a range of behavioral tasks .
As a task that most robots need to be able to complete, navigation is a commonly-studied higher level behavior. For example, Ken Stanley and colleagues have often used maze-solving tasks for simulated robots as an illustrative example problem . Navigation in a physical robot has also been used as proof-of-concept task for demonstrating that algorithms evolved in Avida can successfully be translated to physical robots .
Soft-bodied robots are a particularly popular substrate for research on the evolution of robot morphology . These robots are traditionally evolved in a computer as a cube composed of voxels made up of various different materials with different properties. Recently, this approach has been translated to a non-simulated context via the xenobots project .