Three Generations of Automatically Designed Robots

1
Three Generations of Automatically Designed Robots

• Jordan B. Pollack, Hod Lipson, Gregory Hornby,
  Pablo Funes

2
The Economics of Robots

• High costs associated with robot design,
  construction, and control have frozen their
  development
• Robots in industry are only applied to simple and
  highly repetitive manufacturing tasks (painting,
  arc welding, etc.)

3
Economics of Robotics Continued

• More sophisticated machines (which utilize
  sensors and actuators have found important
  applications) yet decision making has yet to be
  left to the on-board software to any significant
  extent

4
Of Central Importance

• The authors address the issue of how to achieve a
  higher level of complex physicality under control
  at a lower cost
• They are seeking to develop (or facilitate
  further development) of a greater amount of
  control, more sensors, and more nonlinear
  interacting degrees of freedom without incurring
  huge fixed costs
• They believe that this can only be achieved once
  robot design and construction are fully automatic

5
Automatic Construction and Design

• Traditionally, robots are designed on a hardware
  first, software last basis
• Pollack et al. take a more evolutionary
  approach to robot construction
• They believe that the costs of building and
  programming robots will remain so high that the
  direct engineering of robots will continue to be
  an economic failure

6
The Chicken or the Egg?

• In nature, the body and mind of an organism are
  tightly coupled and this is achieved via a series
  of small mutual adaptations
• Robots, like organisms, require a sophisticated
  integration between brain, physicality and
  environment
• Authors have been attempting to achieve
• co-evolution of brain and body within
• robots

7
Automatically Designed Machines of the Future

• The authors see three emergent technologies which
  may make possible a new industry of inexpensive
  automatically designed robots
• Increasing fidelity of advanced mechanical
  simulation (CAD)
• Rapid, one-off prototyping and manufacture
• Understanding of the dynamics of coevolutionary
  learning in the self-organization of complex
  systems

8
Robot Evolution?

• Coevolution creates a series of learning
  environments of increasing complexity and a
  series of learners which are tuned to adapt in
  those environments
• The authors are ultimately concerned with the
  creation of mechanical structures that have
  complex physicality and more degrees of freedom
  than anything that has ever been controlled by
  human designed algorithms
• These new robots would have lower engineering
  costs than currently possible because of minimal
  human design involvement

9
Generation 1 Legobots

• Features for a simulator for evolving morphology
  are
• Representation
• Conservative
• Efficient
• Buildable
• The authors created a simulator for Lego bricks
  that used simple evolutionary algorithms to
  create complex Lego structures
• The algorithm reliably builds structures that
  meet simple fitness goals, exploiting the
  physical properties implicit in the simulation

10
Generation 2 GOLEM

• Robot morphology constrained to be buildable by a
  commercial off-the-shelf rapid prototyping
  machine
• Robots composed of only linear actuators and
  sigmoidal control neurons encased in an arbitrary
  truss-like thermoplastic body
• Entire configuration evolved for a particular
  task and selected individuals were printed
  preassembled

11
GOLEM Evolutionary Process

• Simple steady-state genetic algorithm with
  fitness-proportionate selection and random
  replacement over a population of 200 robots that
  were initially comprised of zero bars and zero
  neurons
• Evolutionary process typically continued for
  300-600 generations and both morphology and
  control evolved simultaneously until it converged
  on creatures that are sufficiently proficient at
  the given task
• Locomotion
• Lifting
• Reaching

12
GOLEM Evolutionary Process

• For locomotion, fitness was awarded according to
  the absolute distance traveled over a specified
  period of neural activation
• The evolved robots exhibited several forms of
  locomotion
• Crawling (virtual, physical)
• Ratcheting (virtual, physical)
• Pedalism

13
Generation 3 Modularity Generative Design
(Tinkerbots)

• Addresses the issue of whether evolutionary
  automatic design techniques can attain the higher
  level of complexity necessary for practical
  engineering projects
• Generative specification, which is an alternative
  to direct encoding, is a grammatical encoding
  that specifies how to construct a design
• Instead of evolving the structure, we evolve a
  set of rules that in turn generate the structure

14
Modularity Generative Design

• A fully automated design system would ideally
  start with a library of basic parts and would
  create new and increasingly complex modules from
  those in its library
• Modularity promotes decoupling and reduces
  complexity
• A design built with a set of standard modules is
  more robust and adaptable, and enhances field
  repair as opposed to a design in which every
  component is unique

15
Modularity Generative Design

• These robots focus on modular design and use
  L-systems as the genotype evolved by the
  evolutionary algorithm
• L-systems are a grammatical rewriting system
  introduced by Lindenmeyer in 1968 to model the
  biological development of multicellular organisms
• Rules are applied in parallel to all characters
  in the string just as cell divisions happen in
  parallel in multicellular organisms

16
Conclusion

• Can evolutionary and coevolutionary techniques be
  used in the design of real robots as Aritificial
  Lifeforms?
• Limitations
• These machines do not have sensors and are not
  really interacting with their environments
• Feedback from how robots perform in the real
  world is not fed back into simulations
• How complex can a simulated system be before the
  errors generated by transfer to reality are
  overwhelming