![]()
Visiting guest. Why not sign in?
|
|
Bipedial walking robots | |
Bipedial walking robots
Hi all, this is my first post. I've had a bit of a dig around on the net & this site looking for an introduction to algorithms to do with bipedial motion. I have a reasonably accurate physics system & now I want to create a bipedial agent to wander about the world. Does anyone have any good links that will increase my knowledge & give me an idea on how to get a virtual pet wandering around without my supervision. Cheers |
|
Physics and AI
Welcome to the boards bradbeveridge. Speaking of skeletons, Andy forwarded me this really cool link: http://vectorlounge.com/04_amsterdam/jam/wireframe.html Basically, it seems like a set of springs. There's some Inverse Kinematics going on there, but it doesn't seem to advanced -- the body can end up in very dodgy positions sometimes, though it does well most of the time. This applet shows how you can get something walking with just emergent physics. You drag the bar at the top, and it walks by trying to satisfy minimum constraints. Don't ask me for more details, I'm not sure! A friend of mine is also doing a 3rd year project on learning motion. It's tough business, and I've no idea how far he's got -- it was a very ambitious project. You can train up neural network controllers to do this, using evolution of somekind. The key will be deciding what inputs/outputs to give your system; most likely the inputs will be the relative rotation of each limb, and the outputs will be the desired force. Your fitness function will also be very important, since getting it wrong will just get weird behaviour. Another friend who did something similar a few years ago managed to get it to learn to collapse forwards while jumping: his fitness function was distance covered... not quite walking, but cool to watch anyway! Let me know if you need more info. It sounds like a great project, something which I want to tackle too at some stage ;) And make sure to let us know how it goes!! |
|
Legged movement
Hi Brad, A while ago I stumbled across this website, which is a demonstration of virtual robots that had control over their own structure via self-evolution. http://www.demo.cs.brandeis.edu/golem/ With only knowledge of the possible joints they allowed to use, these virtual robots (of which they actually built the best performing) pretty quickly learnt to 'walk' Most of the robots wouldnt use what you'd call normal motion though. The following pdf is a syllabus for a course on robotic legged motion, http://www.ai.mit.edu/projects/leglab/6894/handouts/syllabus.pdf good luck with the little dude, let us know when he/she can walk Andy |
|
thanks
Thanks for the links guys, some great sites there - especially the golem project. I'm going to play around with neural nets tonight to try and solve some simple problems, it should be fun. I'm thinking that I might want to try and use a genetic algorithm to create my neural net - has anyone done this & had success. I initially came to this site because I wanted to solve my walking problem without hand programming motion control & now I'm very interested in this whole AI thing, it is really quite cool! Brad |
|
Karl Sims
Yeah, AI can get really funky before you even realise it ;) That can be a problem in the long run, but my suggestion is to just enjoy it :P If you're going to look into evolution, Karl Sims' work is a landmark. See this thread for more information. |
|