evolutionary robotics the italian approach the khepera robot (1996) developed at epfl lausanne,...
TRANSCRIPT
Evolutionary Robotics
The Italian ApproachThe Khepera robot (1996)
Developed at EPFL Lausanne, Switzerland(!)by Francesco MondadaDiameter: 55 mm
Could perform experimentson a table, rather than in alarge arena.
Modular architecture:
Evolutionary Robotics
The Italian ApproachThe Khepera robot (1996)
Developed at EPFL Lausanne, Switzerland(!)by Francesco MondadaDiameter: 55 mm
Could perform experimentson a table, rather than in alarge arena.
Modular architecture:
Evolutionary Robotics
The Italian Approach
Doing evolutionary roboticson a physical robot.
Robot tethered to computer:
computer does the evolution;
robot does the behaving.
Laser-emitting deviceEmits lasers into the workspace
Khepera has a positioningturret attached:
can detect the lasers andthus compute its position (x,y)and heading ().
What behaviors could evolve?
Evolutionary Robotics
The Italian Approach
Evolution of simple navigation:
Evolve a neural network such thatthe robot…
1. Circles through maze as fast as possible.
2. Does not hit the walls.
Robot is given the NN architecture shown.
Robot’s wheels can…rotate backwards quickly (-0.5)stay still (0.0)rotate forward quickly (0.5)or anything in this range (-0.5,0.5)
Robot’s eight proximity sensors return…0 if obstacle is 5cm or further (or absent)1 if proximity sensor is touching object.
Evolutionary Robotics
The Italian Approach
Evolution of simple navigation:
Evolve a neural network such thatthe robot…
1. Circles through maze as fast as possible.
2. Does not hit the walls.
Robot is given the NN architecture shown.
Robot’s wheels can…rotate backwards quickly (-0.5)stay still (0.0)rotate forward quickly (0.5)or anything in this range (-0.5,0.5)
Robot’s eight proximity sensors return…0 if obstacle is 5cm or further (or absent)1 if proximity sensor is touching object.
Create a fitness functionto select for this behavior:
=
Hints:
=0 for worst performance=1 for best possible performance (may not be reachable) vL,vR = speed of left/right wheel i1,…i8=value of proximity sensor
Evolutionary Robotics
The Italian Approach
Evolution of simple navigation:
Evolve a neural network such thatthe robot…
1. Circles through maze as fast as possible.
2. Does not hit the walls.
Robot is given the NN architecture shown.
Robot’s wheels can…rotate backwards quickly (-0.5)stay still (0.0)rotate forward quickly (0.5)or anything in this range (-0.5,0.5)
Robot’s eight proximity sensors return…0 if obstacle is 5cm or further (or absent)1 if proximity sensor is touching object.
Create a fitness functionto select for this behavior:
=
Evolutionary Robotics
The Italian Approach
A sample run:
One generation:40 minutes
100 generations: 66 hours2.77 days
Line segment:
Center of segmentIndicates robot’s position
Line passes through therobot’s left and right wheels.
Evolutionary Robotics
The Italian Approach
Fitness componentsfor the best controller ateach generation.
Q: How did the robot’sbehavior change over time?
Evolutionary Robotics
The Italian Approach
How the fitness componentschanged for the best evolved controller,as it controlled the robot (~1 minute)
How the fitness componentschanged over evolutionary time(over 2.77 days)
Evolutionary Robotics
The Italian Approach
If synaptic weights(thickness of lines) arebilaterally symmetric,
robot gets stuck in corners.
Why?
Evolved robot that approachescorners does not get stuck.
Why do you think this is so?
Evolved controllers onlyever drove the robot at aMaximum speed of 48 mm/s.
Actual top speed is 80 mm/s.
Why not drive at top speed?
The English Approach
Gantry Robot (University of Sussex, Inman Harvey et al., 1994).
Three ways to move (degrees of freedom); three motors.
Evolutionary Robotics
The English Approach
Evolutionary Robotics
Gantry Robot (University of Sussex, Inman Harvey et al., 1994).
Seven ways to sense:
The English Approach
Incremental evolution:
1. approach back wallfrom 4 different initial positions
… evolve …
… success!
2. approach back rectanglefrom 4 different initial positions
…evolve …
… success!
3. avoid back rectangle, approach triangle.from 4 different initial positions
1 =
2 =
3 =
Evolutionary Robotics
The English Approach
An evolved solution:
Robot does not usebump sensors;
Only uses 2 of the 3visual fields.
Q: How could it discriminatebetween the two shapeswithout a ‘recognize triangle’or ‘recognize rectangle’algorithm?
Evolutionary Robotics