[ODE] Re: Virtual creatures (karl sims etc...)

[Gene Ruebsamen]

Hello,

I've been away for a few weeks, and am catching up on this thread.
I'm amazed and happy to hear that there are other's working on this same
problem.  For the past few months I've been using ODE in conjunction with a
GA/ANN combination to "evolve" virtual creatures.  So far, my evolution
consists of only evolving the "brain" (ANN) of the creature and not the
morphology; however, I've had some very interesting results thus far.

I too have had serious problems with joint interpenetration, and the
creatures taking advantage of "errors" in the simulation.  When I get a
joint error, often the creature will spasm and explode.  Since my fitness
function is based mostly upon the distance travelled by the creature (im
trying to evolve locomotion), when a creature explodes it is generally
assigned a high fitness value. Thus my entire simulation, over time, tends
toward instability.

I've managed to "constrain" the fitness function a bit, and solve the above
problems; however, I have also had to increase the number of calculations
per step to achieve stability... So far, I've had promising results.  I've
created several different creatures, and most of them have learned to "walk"
so far.  I'll try and get some MPEG's up soon.

One thing I would like to see is the implementation of the
dJointGetAMotorAngleRate() function.  As it stands now, I cannot feed back
velocity information from any AMotor joints because that function hasn't
been implemented yet.  I've had better luck feeding back velocity
information to the Neural Network vs. static positional information.
Probably because the ANN does not have to integrate the data to compute
velocity, thus there is one less step for the ANN to perform.

Gene Ruebsamen

> -----Original Message-----
> From: ode-admin@q12.org [mailto:ode-admin@q12.org]On Behalf Of Ian
> Sherratt (Shez)
> Sent: Friday, May 17, 2002 7:45 AM
> To: ode@q12.org
> Subject: Re: [ODE] Re: Virtual creatures (karl sims etc...)
>
>
> A simple, nasty way to fix this particular problem may be to
> simulate your
> entire population at a fairly high timestep/low accuracy, take the top x%
> that you are considering for breeding, and simulate those under the same
> conditions at a higher accuracy.  This only works for 'classical' GA/GP
> systems where there is either very simple or no interaction between the
> evolving creatures.
>
> Just a thought :-)
> Cya
> Shez
> _______________________________________________
> ODE mailing list
> ODE@q12.org
> http://q12.org/mailman/listinfo/ode
>