Boston Chapter, IEEE Robotics and Automation Society Meeting

                    Tuesday, 10 October, 2000
       Informal Discussion: 6:00 P.M.; Program: 6:30 P.M.
              Natural MicroSystems, Framingham, MA


                    CO-EVOLUTIONARY ROBOTICS

                Hod Lipson and Jordan B. Pollack
                   Computer Science Department
                       Brandeis University


                            ABSTRACT

   Can evolutionary principles be used to automate design of
robot morphology and control? Can bodies and brains co-evolve
together as they did in nature, stimulating and constraining each
other, to yield new machines which are not laboriously designed
by human engineers?

   In our lab, we work on the evolution of real robots based on
abstract biological principles: neural networks, which simulate
mathematical principles thought to be involved in brains;
evolutionary computation, which uses software to simulate
Darwinian evolution; and co-evolution, which captures the
principles of "arms-races" which arises among multiple
cooperating and competing agents.

   We will show some robots who evolved inside a simulated world
of Lego blocks and were built by hand. These robots did not move,
but demonstrated that the interaction of evolution and physics
can lead to primitive forms of discovery, and what can be evolved
in simulation can be transferred to reality.

   We will also demonstrate robots who evolved arbitrary
morphologies and neural controllers for locomotion inside a
quasi-dynamic motion simulator and then were converted to reality
using rapid prototyping machinery. The robot is born as a solid
three-dimensional structure without the need for tooling or human
intervention. Support tissue is disposed, motors are then snapped
in, and the evolved neural network is downloaded to a
microprocessor. The evolved creatures worked in reality as they
did in simulation.

   This work establishes for the first time robotic replication,
a complete physical evolution cycle, where a robotic system can
design and manufacture new robots. However, it is not yet
self-replication, because the machines produced are not as
capable as their forbears.


                       ABOUT THE SPEAKERS

   Hod Lipson completed his Ph.D. in mechanical engineering at
the Techion - Israel Institute of Technology in 1998, after
serving five years as the head of the CAD/CAM R&D office in the
Israeli Navy. Hod Lipson is now a research scientist at the
Dynamical Evolutionary Machine Organization (DEMO) Lab at
Brandeis University and at the mechanical engineering department
at MIT. He is interested in the area of design automation.

   Jordan Pollack received the Ph.D. in computer science in 1987
from the University of Illinois and is on the faculty of Brandeis
University. His interests span artificial intelligence,
artificial life, complex systems, neural computation,
evolutionary computation, co-evolution, robotics, games, and
educational technology.