Automatic design and Manufacture of Robotic Lifeforms

The field of Artificial Life examines “life as it could be” based on understanding the principles and simulating the mechanisms of real biological forms. Just as airplanes use the same principles as birds, but have fixed wings, artificial lifeforms may share the same principles, but not the same implementation in chemistry. Every feature of living systems seems wondrous until it is understood: Stored energy, autonomous movement, and even animal communication are no longer miracles, as they are replicated in toys using batteries, motors, and computer chips.

The field of Artificial Life examines “life as it could be” based on understanding the principles and simulating the mechanisms of real biological forms. Just as airplanes use the same principles as birds, but have fixed wings, artificial lifeforms may share the same principles, but not the same implementation in chemistry. Every feature of living systems seems wondrous until it is understood: Stored energy, autonomous movement, and even animal communication are no longer miracles, as they are replicated in toys using batteries, motors, and computer chips.

Complex biological forms reproduce by taking advantage of an arbitrarily
complex set of auto-catalyzing chemical reactions. Biological life is in control of its
own means of reproduction, and this autonomy of design and manufacture is a key
element which has not yet been understood or reproduced artificially. To this date,
robots – a form of artificial life – are still designed laboriously and constructed by
teams of human engineers at great cost. Few robots are available because these costs
must be absorbed through mass production that is justified only for toys, weapons,
and industrial systems like automatic teller machines.

Here we report a set of experiments in which simple electro-mechanical systems
evolve from scratch to yield physical locomoting machines. Like biological lifeforms
whose structure and function exploit the behaviors afforded by their own chemical
and mechanical medium, our evolved creatures take advantage of the nature of
their own medium – thermoplastic, motors, and artificial neurons. We thus achieve
autonomy of design and construction using evolution in a limited universe physical
simulation coupled to off-the-shelf rapid manufacturing technology. This is the
first time robots have been robotically designed and robotically fabricated.

The field of Artificial Life examines “life as it could be” based on understanding
the principles and simulating the mechanisms of real biological forms. Just as airplanes
use the same principles as birds, but have fixed wings, artificial lifeforms may share the
same principles, but not the same implementation in chemistry. Every feature of living
systems seems wondrous until it is understood: Stored energy, autonomous movement,
and even animal communication are no longer miracles, as they are replicated in toys
using batteries, motors, and computer chips.

Our key claim is that to realize artificial life, full autonomy must be attained not
only at the level of power and behavior (the goal of robotics, today), but also at the
levels of design and fabrication. Only then can we expect synthetic creatures to bootstrap and sustain their own evolution. We thus seek automatically designed and constructed
physical artifacts that are (a) functional in the real world, (b) diverse in architecture
(possibly each slightly different), and (c) producible in short turn-around time, low cost
and large quantities. So far these requirements have not been met.

The experiments described here use evolutionary computation for design, and
additive fabrication for reproduction. The evolutionary process operates on a population
of candidate robots, each composed of some repertoire of building blocks. The
evolutionary process iteratively selects fitter machines, creates offspring by adding,
modifying and removing building blocks using a set of operators, and replaces them into
the population (see methods section). Evolutionary computation has been applied to many
engineering problems. However, studies in the field of evolutionary robotics reported
to date involve either entirely virtual worlds or, when applied in reality, adaptation of
only the control level of manually designed and constructed robots. These robots
have a predominantly fixed architecture, although Lund evolved partial aspects of the
morphology, Thompson evolved physical electric circuits for control only, and we
evolved static Lego structures, but had to manually construct the resultant designs.
Other works involving real robots make use of high-level building blocks comprising
significant pre-programmed knowledge. Similarly, additive fabrication technology has
been developing in terms of materials and mechanical fidelity but has not been placed
under the control of an evolutionary process.

Our approach is based on use of only elementary building blocks and operators in
both the design and fabrication process. As building blocks are more elementary, any
inductive bias associated with them is minimized, and at the same time architectural
flexibility is maximized. Similarly, use of elementary building blocks in the fabrication
process allows it to be more systematic and versatile. As a theoretic extreme, if we could
use only atoms as building blocks, laws of physics as constraints and nano-manipulation
for fabrication, the versatility of the design space would be maximized. Earlier reported
work used higher-level components and limited architectures (like only tree structures)
resulted in expedited convergence to acceptable solutions, but at the expense of
truncating the design space. Furthermore, these design spaces did not consider

The design space we used was comprised of bars and actuators as building blocks
of structure and artificial neurons as building blocks of control. Bars connected with free
joints can potentially form trusses that represent arbitrary rigid, flexible and articulated
structures as well as multiple detached structures, and emulate revolute, linear and planar
joints at various levels of hierarchy. Similarly, sigmoidal neurons can connect to create
arbitrary control architectures such as feed-forward and recurrent nets, state machines and
multiple independent controllers (like multiple ganglia). Additive fabrication, where
structure is generated layer by layer, allows automatic generation of arbitrarily complex
physical structures and series of physically different bodies, including any composed of
our building blocks. The
bars connect to each other through ball-and-socket joints, neurons can connect to other
neurons through synaptic connections, and neurons can connect to bars. In the latter case, the length of the bar is governed by the output of the neuron, by means of a linear
actuator. No sensors were used.

Starting with a population of 200 machines that were comprised initially of zero
bars and zero neurons, we conducted evolution in simulation. The fitness of a machine
was determined by its locomotion ability: the net distance its center of mass moved on an
infinite plane in a fixed duration. The process iteratively selected fitter machines, created
offspring by adding, modifying and removing building blocks, and replaced them into the
population. This process typically continued for 300 to 600 generations. Both body (morphology) and brain (control) were thus co-evolved

The simulator we used for evaluating fitness (see methods section) supported
quasi-static motion in which each frame is statically stable. This kind of motion is
simpler to transfer reliably into reality, yet is rich enough to support low-momentum
locomotion. Typically, several tens of generations passed before the first movement
occurred. For example, at a minimum, a neural network generating varying output must
assemble and connect to an actuator for any motion at all. Various patterns of evolutionary dynamics emerged, some of which are
reminiscent of natural phylogenic trees.

Selected robots out of those with winning performance were then automatically
replicated into reality: their bodies, which exist only as points and lines, were first
converted into a solid model with ball-joints and accommodations for linear motors
according to the evolved design. This solidifying stage was performed by a
automatic program which combined pre-designed components describing a generic bar,
ball joint, and actuator. The virtual solid bodies were then materialized using commercial
rapid prototyping technology. This machine used a temperature-controlled head
to extrude thermoplastic material layer by layer, so that the arbitrarily evolved
morphology emerged as a solid three-dimensional structure without tooling or human
intervention. The entire pre-assembled machine was printed as a single unit, with fine plastic supports connecting between moving parts; these supports broke away at
first motion. The resulting structures contained complex joints that would be difficult to
design or manufacture using traditional methods. Standard stepper
motors were then snapped in, and the evolved neural network was executed on a
microcontroller to activate the motors. The physical machines (3 to date) then faithfully
reproduced their virtual ancestors’ behavior in reality.

In spite of the relatively simple task and environment (locomotion over an infinite
horizontal plane), surprisingly different and elaborate solutions were evolved. Machines
typically contained around 20 building blocks, sometimes with significant redundancy
(perhaps to make mutation less likely to be catastrophic). Not less surprising was the
fact that some exhibited symmetry, which was neither specified nor
rewarded for anywhere in the code; a possible explanation is that symmetric machines are
more likely to move in a straight line, consequently covering a greater net distance and
acquiring more fitness. Similarly, successful designs appear to be robust in the sense that
changes to bar lengths would not significantly hamper their mobility. Three samples are
shown and described in detail in Figure 5, exploiting principles of ratcheting (5a), anti-phase
synchronization (5b) and dragging (5c). Others (not shown here) used a sort of a
crawling bi-pedalism, where a body resting on the floor is advanced using alternating
thrusts of left and right “limbs”. Some mechanisms use sliding articulated components to
produce crab-like sideways motion. Other machines used a balancing mechanism to shift
friction point from side to side and advance by oscillatory motion. Table I compares the
performances of three physical creatures to their virtual ancestors. Note that although
overall distance traveled in the 2nd and 3rd cases does not match, in all cases the physical motion was achieved using corresponding mechanical and control implementations. The
difference in distance results from slipping of the limbs on the surface, implying that the
friction model used in the simulation was not realistic.

In summary, while both the machines and task we describe in this work are fairly
simple from the perspective of what human teams of engineers can produce, and what
biological evolution has produced, we have demonstrated for the first time a robotic
bootstrap, where automatically designed electromechanical systems have been
manufactured robotically. We have carefully minimized human intervention both in the
design and in the fabrication stages. Besides snapping in the motors, the only human
work was in informing the simulation about the universe that could be manufactured.

Without reference to specific organic chemistry, life is an autonomous design
process that is in control of a complex set of chemical factories allowing the generation
and testing of physical entities which exploit the properties of the medium of their own
construction. Using a different medium, namely off-the-shelf rapid manufacturing, and
evolutionary design in simulation, we have made progress towards replicating this
autonomy of design and manufacture. This is the first time any artificial evolution system
has been connected to an automatic physical construction system. All together, our
evolutionary design system, solidification process, and rapid prototyping machine form a
primitive “replicating” robot. While there are many, many further steps before this
technology is dangerous, we believe that if indeed artificial systems are to ultimately
interact and integrate with reality, they cannot remain virtual; it is crucial that they cross
the simulation-reality gap to learn, evolve and affect the physical world directly.
Eventually, the evolutionary process must accept feedback from the live performance of
its products.

Future work is primarily needed in understanding how more complex modular
structures might self-organize, and how these complex structures may transfer into reality
under control of the evolutionary process. Technological advances in MEMS, nano-fabrication and multi-material rapid prototyping that can embed circuits and actuators in bulk structure, higher fidelity physical simulation, and increased understanding of
evolutionary computational processes may pave the way for what Moravec has termed
“Escape Velocity”.


Robot representation: A robot is represented by a string of integers and floating point
numbers that describe bars, neurons, and their connectivity, as follows:

robot := (vertices)(bars)(neurons)(actuators)

vertex := (x,y,z)

bar := (vertex1 index, vertex2 index, relaxed length, stiffness)

neuron := (threshold, synapse coefficients of connections to all neurons)

actuator := (bar index, neuron index, bar range)

Evolution process: Experiments were performed using version 1.2 of GOLEM
(Genetically Organized Lifelike Electro Mechanics), which can be obtained from We carried out a simulated evolutionary
process: The fitness function was defined as the net Euclidean distance that the center-of-mass
of an individual has moved over a fixed number of (12) cycles of its neural control.
We started with a population of 200 null (empty) individuals. Each experiment used a
different random seed. Individuals were then selected, mutated, and replaced into the
population in steady-state as follows: The selection functions we tried were random,
fitness proportionate or rank proportionate. The mutation operators used to generate an
offspring were the following (with probability): Small mutation in length of bar or
neuron synaptic weight (0.1), removal/addition of a small dangling bar or unconnected
neuron (0.01), split vertex into two and add a small bar or split bar into two and add
vertex (0.03), attach/detach neuron to bar (0.03). The dice were rolled until at least one
mutation was applied. The mutations were applied directly to the symbolic representation
of the phenotype. After mutation, a new fitness is assigned to the individual by means of
a simulation of the mechanics and the control. The offspring was
inserted into the population by replacing an existing individual. The replacement functions we tried chose individuals to replace either randomly, in inverse-proportion to
its fitness, or using similarity proportionate criteria (deterministic crowding). Various
permutations of selection-replacement methods are possible; The results we report here
were obtained using fitness proportionate selection and random replacement. However,
using rank selection instead of fitness proportionate selection, or using random selection
with fitness proportionate replacement yields equivalent results. . The process continued
for 500-5000 generations (approx 10 to 10 evaluations overall). The process was carried
out both serially and in parallel (on a 16-processor computer). On parallel computers we
noticed an inherent bias towards simplicity: Simpler machines could complete their
evaluation sooner and consequently reproduce more quickly than complex machines (this
could be avoided with a generational implementation).

Our evolutionary simulation was based on Evolutionary Srategies and
Evolutionary Programming, since it directly manipulated continuous valued
representations and used only elementary operators of mutation. Alternatively, we could
have used Genetic Algorithms and Genetic Programming that introduce cross-over
operators sensitive to the structure of the machines, which might change the rate of
evolution and lead to replicated structures. We did not form a morphological grammar
from which the body is developed, but evolved directly on the symbolic representation
of the phenotype. And, instead of separating body (morphology) and brain (control) into
separate populations, or providing for a “neonatal” stage that might allow us to select for
brains that are able to learn to control their bodies, we simply applied selection to bodies
and brains as integrated units. This simplified experimental setup followed our focus on
completing the simulation and reality loop, but we anticipate that the many techniques
that have been developed in evolutionary and co-evolutionary learning will enrich
our results.

Simulation: Both the mechanics and the neural control of a machine were simulated
concurrently. The mechanics were simulated using quasi-static motion, where each frame
of the motion was assumed to be statically stable. This kind of motion is simple to
simulate and easy to induce in reality, yet is rich enough to support various kinds of low-momentum
motion like crawling and walking (but not jumping). The model consisted of ball-joined cylindrical bars with true diameters. Each frame was solved by relaxation: An
energy term was defined, taking into account elasticity of the bars, potential gravitational
energy, and penetration energy of collision and contact. The degrees of freedom of the
model (vertex coordinates) were then adjusted iteratively according to their derivatives to
minimize the energy term, and the energy was recalculated. Static friction was also
modeled. The use of relaxation permitted handling singularities (e.g. snap-through
buckling) and under-constrained cases (like a dangling bar). Noise was added to ensure
the system does not converge to unstable equilibrium points, and to cover the simulation-reality
gap. The material properties modeled correspond to the properties of the rapid
prototyping material (E=0.896GPa, ñ =1000Kg/m 3
ó yield=19MPa). The neural network
was simulated in discrete cycles. In each cycle, actuator lengths were modified in small
increments not larger than 1 cm.

Author: Hod Lipson and Jordan B. Pollack

News Service: TheExperiment Network


Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.