Autonomous Embodied Evolution of a Biomimetic Robot's Rhythmic Motion Behavior
Speeding up the recorded movements
digitally underlines the individual characteristics of the repetitive rhythmic
Pattern 58 of evolutionary
Pattern 71 of evolutionary
Pattern 124 of evolutionary
Pattern 186 of evolutionary
of patterns with the highest performance values over the course of the evolutionary
run 19_frog. Each pattern's fitness results are listed in consecutive
Several evolutionary runs have been conducted on two
different robotic creatures. By focusing the evolutionary computation on only
a few springs clear results were achieved. Then the evolution was able to
successfully discover the springs' individual dynamics and modify the actuation
A steady rising fitness value during the ongoing
The fluctuations in hardware dynamics and environmental conditions that were
meant to supply the unstable fitness landscape were mainly caused by defective
electronic connections. Several springs tended to become inactive during long
24-hour-runs. This instability was not intended yet served as good steering
material for the evolutionary computation. If the fitness weight values of
the evaluation function were set accordingly, the evolution was able to detect
which springs were inactive and to compensate for these errors.
The evolution copes with the loss of springs by reducing their heating time to prevent waste of energy.
The two-dimensional perception of the monitoring video camera and a mere focus
on simple motion detection doesn't allow identification of individual joint
movements. The overall motion of the whole creature is summed together and
this makes it impossible to distinguish if the resulting motion originates
from active or passive relocation. The more springs were employed the more
complex the cause-and-action relationship of activation and movement became
and the less obvious trends could be recognized in pattern design. This missing
talent of coping with higher complexity can maybe be explained with the too
small population size, the too high mutation rates and the too short evolutionary
By taking the human designer out of the loop, the
artwork shows the shift from human constraints to machine constraints.
The pure information about the robot's movements is filtered and distorted
by the eye of the camera. This imperfection leads to behaviors that can be
perceived as cheating, when the evolution presents solutions that
might satisfy the program's fitness evaluation yet not the human inspector.
Here the system reaches a certain degree of autonomy from the human intentions
that guided the design of the evolutionary framework.
Departing from its foreseen path the system escapes
into its own world of logic.
The two-dimensional view of the camera excluded for example movements that
acted in parallel direction with the viewing angle of the lens. This made
the evolution disregard the importance of certain joint movements. If the
influence of the energy usage was high enough in the fitness evaluation, the
evolution was partially able to discover this waste of energy. By perceiving
the particular spring elements as less influental in the motion production
or even inactiv, the system reacted by reducing the specific spring heating
times as described before. Yet this effect was only possible if the number
of active springs were kept low.
Even though most evolutions didn't necessarily produce highly representable
results in terms of evolutionary strategies, they still succeeded in creating
different rhythmic motion behaviors. In the real-time exhibition setup the
movements of the robotic creatures emphasize more the slowness and noiselessness
of the SMA springs, whereas speeding up the recorded movements digitally underlines
the individual characteristics of the repetitive rhythmic actions. Displaying
cycle times of 60 seconds as a 3 seconds loop, transforms the former abstract
limb actuations into associative flapping, hopping or walking behaviors. The
time lapse videos allow to recognize the fine distinctions between the individual
patterns much better and create much higher mimesis effects in the mind of
the viewer. The creatures suddenly become animals performing locomotion techniques
in free air and draw associative character traits like jerkiness or jumpiness.
For more detailed results, download my papers ...
(c) 2007 - Thesis project by Eva Schindling
Completed in the MSc. programme Art & Technology