Real-Time Physics Simulation Forum

This paper is awesome -- two simple ideas that combine to produce some pretty amazing results: http://www.cs.ubc.ca/~van/papers/simbicon/simbicon.html

Those recently interested by "Sumotori" should check out the videos as well.

raigan

raigan2 wrote:This paper is awesome -- two simple ideas that combine to produce some pretty amazing results: http://www.cs.ubc.ca/~van/papers/simbicon/simbicon.html

Those recently interested by "Sumotori" should check out the videos as well.

raigan

really nice results, although the ideas look a bit more than two(FEL,Fourier,Motion Graph, Balancing strategy etc...). I wonder why interactive demos are never released, it would be cool to play with it.

cheers.
Antonio

I guess I was forgetting about the parts I wouldn't be attempting, like Fourier/keyframe-tracking and FEL

I think I'm missing something about FEL: it seems like what they're doing is procedurally generating a canned animation (where dynamic values -- torques -- are animated instead of kinematic values). This animation is used to give the controller some information about the expected future state..

"Playback of prerecorded torques" seems like a weird way to predict the future -- if you really want to use "feedback from the future" in your controllers, and you're simulating everything, it seems like a better idea would be to take a big/approximate step forward and use that as the predicted future state.

Then again, that wouldn't work for a real robot.

My main concern with the "learned model" feedforward controller is that if the landscape changes, the learned model will be invalid.

raigan