Ragdoll physics without 3rd party libraries

[Superposed_Cat]

Hey all, are there any resources available online on rag doll physics programming without 3rd party libraries? the algorithms etc? Any help appreciated.

 

[Greg Bernhardt]

Here are a few to check out
http://www.cs.rpi.edu/~cutler/classes/advancedgraphics/S07/final_projects/mulley_bittarelli.pdf
http://gamedevelopment.tutsplus.com...s-with-simple-verlet-integration--gamedev-519
http://tomlee.co/misc/jakobsens-advanced-character-physics-annotated/

 

[Greg Bernhardt]

What kind of project are you working on?

 

[Superposed_Cat]

What kind of project are you working on?

Make a simulated robot learn to walk with genetic algorithms. I wanted to do it with neural networks but couldn't figure out what to make the inputs and outputs

 

[kroni]

Ones you have a simulator, i personally think that it's not a good idea to not use a third party library. You will loose a lot of time to build another (and another) body simulator. I think that you need to concentrate on the interesting part. For the input and output it depend on the library, you can use joint angle and body position and the neural net will give in output the joint momentum, or the joint angle speed to control the caracter.

Previously i try genetic algorithm for this kind of task and it doesn't work well because learning is very long. Walk can be learned by logic, automatic. You can automaticaly learn physics law with généric model regression like least square regression and then use it to search how to move using a tree to explore the configuration space.

 

[Superposed_Cat]

Previously i try genetic algorithm for this kind of task and it doesn't work well because learning is very long. Walk can be learned by logic, automatic. You can automaticaly learn physics law with généric model regression like least square regression and then use it to search how to move using a tree to explore the configuration space.

Could you elaborate on the least squares technique for learning physics please?

 

[kroni]

Genetic algorithm and neural network are généric function interpolator and regression. To learn physics you can use mathématical tools that fit the observation and make a correlation with the input. So for a value of the input you can get a value for the output, the objective is to learn physical law (what is the consequence of an action). In a second step you can use this model to act on the environement. I know a publication about LWPR, a generic high dimension regression algorithm, and a paper about physics model learning.

Link are :
http://wcms.inf.ed.ac.uk/ipab/slmc/research/lwpr/lwpr-tutorial
http://homepages.inf.ed.ac.uk/svijayak/publications/vijayakumar-ICML2000.pdf

 

[Superposed_Cat]

thanks, out of interest why do you type your 'e's like this? not complaining they are whimsical :P

é

 

[D H]

Hey all, are there any resources available online on rag doll physics programming without 3rd party libraries?

Why do you want to avoid 3rd party libraries? You're using a compiler. Why aren't you using machine language? What's the difference between a 3rd party library and a compiler?

There's a lot to be said for those 3rd party libraries. The underlying math can be very tricky. Properly modeling multi-body physics is trickier yet. The odds you'll get it right are slim to none. There are very good 3rd party libraries that are free, get it right, and are fast.

 

[jtbell]

why do you type your 'e's like this?

If I were a betting man, I would lay odds of at least ten to one that kroni is more comfortable writing in French than in English.

https://www.google.com/?gws_rd=ssl#q=translate+"generic"+to+french

 

[kroni]

I use "é" because i am french and we also use e , é , é , ê , a , à , ç. It seem funny but it's not, it's an boring language.

 

[kroni]

lol