say I want to see if i can get a biped to walk upright with an ANN.
I have a fitness function, and how I plan on adjusting weights is for each output, i test random deviations of the output with the fitness function till i achieve a better result (just ever so slightly just for the next frame), then i back propagate this...
would this work?
Have a look here
Can't find the original video at the moment where they start with a single legged creature and used GA to get it to hop around.
Then they did a biped
wow i must say that truly is amazing, and just to think, it all came from a little laboratory back in 1957 with the first perceptron!
Euphoria is an amazing achievement, but the animations looks just like marionettes. The puppeteering engine needs to be way more proactive to look natural. The most convincing parts were the falling and stumbling, situations where a human would be caught off guard and only instinctively controlling their movement.
The other product they do handles that.
You create networks of animations and it blends between them based on input events.
Ah. The video gave me the impression it was a complete character animation package.
I want to take this idea further!!! Id like a wrestling game or something (take ragdoll gaming to the next level) - it would be really cool.
But setting up the physics system and neural net learning is really challenging.
oh well, i guess god wanted it to be a challenge anyway...
The system mixes generated animations and artist created animations based on events.
It's power comes from the fact that it can blend multiple animations in a realistic way.
The stuff that really makes a difference is the reaction system, for example someone getting shot. Traditionally we have one or two death animations, but with their system it actually has almost an infinite set of reactions depending on the nature of the event.
Makes a big difference.