Evolutionary Robotics

1
Evolutionary Robotics

• Robert Rose
• 5/25/04

2
Agenda

• What is evolutionary robotics?
• Why evolve robots?
• Physical simulation of robots
• Neural networks for robotics
• Evolving robot controllers
• Related work

3
What is Evolutionary Robotics?

• Evolutionary robotics is a new technique for
  automatic creation of autonomous robots inspired
  by the darwinian principle of selective
  reproduction of the fittest. (Nolfi and Floreno)
• How do you evolve a robot?
• Define a fitness function. Jumps highest,
  walks farthest, etc.
• Create thousands of random robots.
• Test their fitness.
• Pick the best robots, mate them to create
  thousands more offspring.
• Go to 3.

4
Why Evolve Robots?

• Goal 1 So youve built a robot. How do you
  control it?
• Many robot behaviors are too difficult to program
  by hand.
• If you could assist the robot by helping it
  learn, you wouldnt need to program it.
• Goal 2 Simulated robots in animation.
• Motion capture and keyframing can not account for
  all possible desired behaviors.
• Combining motion capture with physical simulation
  is a black art that requires much programmer
  intervention and special tuning.

5
Physical Simulation

• Physical simulation models the behavior of
  physical objects in a computer.
• We can simulate a real robot in a computer!
• Physical simulation provides a richer environment
  if our goal is animation.
• Is our goal to evolve a controller for a real
  robot?
• Evolving a robot controller using a physical
  robot is impractical.
• Simulation isnt a perfect representation of the
  physical robot, but we hope its close enough.

6
Neural Networks

• NNs attempt to model the behavior of a group of
  connected neurons.
• A neuron may have one to many inputs.
• A neuron outputs only one value.
• A neuron uses its activation function to
  determine its output value.
• A common activation function is the sigmoid, but
  to quickly develop more complex behavior other
  activation functions can be used sum, product,
  sine, absolute value, greater than, etc. (Sims
  94).

7
Neural Networks for Robotics

• Two common approaches
• Each controller has its own NN
• One big NN with outputs for each controller

8
Evolutionary Robotics

• Based on Genetic Programming
• Define a chromosome format that specifies a
  robot
• Evolve the chromosome by generating many random
  chromosomes, picking the best, mating them,
  repeat.

9
Evolving Robot Controllers

• Define a fitness function
• Create many random controllers (1000)
• Test each controller, rate its fitness
• Take the best controllers (10), mate and
  mutate them to create many more random
  controllers
• Repeat (1000 generations)
• Game Theory can also be used

10
Evolving Virtual Creatures

• Structure of robot itself is evolved
• Karl Sims, Evolving Virtual Creatures, 1994

(Karl Sims genarts.com)
11
Evorobot

• Evolve robot controller for the Khepera robot in
  software.
• Stefano Nolfi
• http//gral.ip.rm.cnr.it/evorobot/simulator.html

K-team.com
12
FOX

• Controller is a neural network
• Gradient descent optimization of controller
• Russell Smith, Intelligent Motion Control with
  Artificial Cerebellum, 1998.

Q12.org
13
Darwin2k

• Software for evolving robot controllers
• www.darwin2k.com

Darwin2k.com
14
Demos!