review articles
Doi: 10.1145/2493883
Taking a biologically inspired approach to
the design of autonomous, adaptive machines.
By Josh c. BonGaRD
Evolutionary
Robotics
the autoMateD Design, construction, and
deployment of autonomous and adaptive machines
an open problem. Industrial robots are an example of
autonomous yet nonadaptive machines: they execute
the same sequence of actions repeatedly. Conversely,
unmanned drones are an example of adaptive yet
non-autonomous machines: they exhibit the adaptive
capabilities of their remote human operators. To date,
the only force known to be capable of producing fully
autonomous as well as adaptive machines is biological
evolution. In the field of evolutionary robotics, 9 one
class of population-based metaheuristics—evolutionary
algorithms—are used to optimize some or all aspects of
an autonomous robot. The use of metaheuristics sets
this subfield of robotics apart from the mainstream
of robotics research, in which machine learning
algorithms are used to optimize the control policya of a
robot. As in other branches of computer science the use
of a metaheuristic algorithm has a cost and a benefit.
The cost is that it is not possible to guarantee if (or
when) an optimal control policy will be found for a given
robot. The benefit is few assumptions must be made
about the problem: evolutionary algorithms can improve both the parameters and the architecture of the robot’s
control policy, and even the shape of
the robot itself.
Because the trial-and-error nature
of evolutionary algorithms requires a
large number of evaluations during
optimization, in many evolutionary
robotics experiments optimization is
first carried out in simulation. Typically an evolutionary algorithm generates populations of virtual robots that
behave within a physics-based simulation.b Each robot is then assigned
a fitness value based on the quality of
its behavior. Robots with low fitness
are deleted while the robots that remain are copied and slightly modified
in some random manner. The new robots are evaluated in the simulator and
assigned a fitness, and this cycle is repeated until some predetermined time
period has elapsed. The most-fit robot
may then be manufactured as a physical machine and deployed to perform
its evolved behavior.
To illustrate the distinction between
mainstream and evolutionary robotics,
consider two experiments drawn from
the two fields. Legged locomotion—
b Interested readers may download and perform
their own evolutionary robotics experiments
at http://www.uvm.edu/~ludobots.
key insights
manual design of a mobile robot
that is autonomous and adaptive is
extremely difficult.
as an alternative, computers can ‘evolve’
populations of robots in a simulator
to exhibit useful behavior and then
manufacture physical versions of the best
ones, very much like how farmers breed
crops for high yield. this approach is
known as evolutionary robotics.
a A control policy is some function that transforms a robot’s sensor signals into
commands sent to its motors.
this evolutionary approach changes
the way we view robotics: rather than
machine-learning techniques improving
behaviors for a hand-designed robot,
focus shifts to creating an evolutionary
system that continuously designs and
manufactures different robots with
increasing abilities.
