Autonomous Embodied Evolution of a Biomimetic Robot's Rhythmic Motion Behavior
The robotic art work Spinal Rhythms
investigates the qualities and dynamics of physical movement performed by
photo credit: akai yoshiyuki
By designing a robot's body as a primitive
abstraction, a connected system of bare limbs linked by articulated joints,
the spotlight lies on the action that brings those inorganic shapes to life
- the motion. It is movement that superimposes life and characteristics onto
the abstract inorganic skeleton.
The movement becomes the expression of the creature's
individualism, the language that communicates its identity.
Actuated by shape memory alloy springs the robot performs slow and elastic
movements that differ from robotics' typical electrical motor characteristics.
Their noiseless activation and stepless transformation grants the robots a
much more graceful and life-like appearance.
The perceived movement of the robots is the interdependent product of software,
hardware and environment. The computation unit sends activation patterns to
the muscles elements, they apply forces onto the mechanics of the robotic
skeleton which are transformed into kinetic energy that moves the whole body
against environmental forces. The close coupling between the control signals
and the mechanical dynamics of the body often makes a clear definition of
the origins of the resulting behavior impossible. The body puts the constraints
on what the control signals are able to do. By performing evolutionary computation
on the actuation patterns, the control system learns to adapt itself to the
morphology it has to deal with. The evaluation process uses image analysis
to grade the performance of motion patterns according to a fixed set of fitness
functions and attempts to find activation patterns that produce more movement
while consuming less energy. The robot gets the chance to experience its own
body through a simple trial and error process and learns to achieve better
and better results with time.
Trained in an autonomous loop without human supervision the robot is granted
a certain awareness of its own body.
Robotic control systems are usually evolved in simulation and are less efficient
than expected due to a lack in complexity in the simulated testing environment.
An embodied evolution solves the reality gap problem by performing
all the tests on the hardware robot itself. The fault-prone hardware-body
of the robot and changing environmental conditions create an unstable fitness
landscape that demands continuous adaptation of the activation patterns. Instead
of training for a fixed set of conditions, the inclusion of possible alteration
during the evolution invokes autonomous adaptation and damage repair in close
interaction with the changing environment.
As the project employs the techniques of artificial evolution in an embodied
setup, it tries to bridge between artificial life art in the digital domain
and in robotics. The crucial differences between digital and analog worlds
- constituted in the messiness and unpredictability of real life - are emphasized
instead of being inhibited. The evolution of movements and motion patterns
has been the subject of projects like Strandbeests
Virtual Creatures, yet the actual evolutionary process has always been
conducted in physical-world simulation programs. Embodied evolutions are yet
rarely carried out in artificial life art and are more common in the evolutionary
robotics discipline. Even though actual reproduction and mutation techniques
are only applied to the control software of the moving robots, the robotic
body and all its constraints and instabilities become the main carrier of
the evolutionary focus.
(c) 2007 - Thesis project by Eva Schindling
Completed in the MSc. programme Art & Technology