|
From: Robert S. <ro...@pr...> - 2006-06-12 14:20:31
|
Vision is a 10 by 10 pixel matrix with 3 channels per pixel and there are 2 eyes so that makes 600 inputs. As for your efficiency question I can't seem to reproduce it... Zachary Bauer wrote: > Hey Rob, > Wow, I am impressed, you've got a lot planned. We'll have to talk > on Friday. A couple quick questions: > > 1.) With the set of brains I have saved and I reload to train, they > are blowing way over 100% efficiency, not sure why. Starting the > simulation from scratch seems to be ok, but I may just need to train > them longer. > 2.) How many vision inputs are we using? Right now vision adds 600 > inputs, so I was wondering if the vision window is bigger or what. > > Anywho, that is all for now. I am making rapid progress on brain > stuff, but it has a LONG ways to go. Somestuff will need to be > tweaked (such as sensor interfaces to the bug/brain) in the new > project, but we can discuss that later. > Welcome back! See ya Friday! > > -Zach > > */Robert Stacey <ro...@pr...>/* wrote: > > Hey guys, > > I'm back from Taiwan. I will be at our friday meeting this week. So > here's a short update. Over the past two weeks I've been playing > around > with a physics engine called ODE, which stands for Open Dynamics > Engine(http://www.ode.org/). This project was written in C, but some > smart folks created a JNI (Java Native Interface) called odejava > (http://odejava.org). The concept here is similar to that of using > JOGL > to access OpenGL libraries. Just like JOGL, there are odejava > binaries > for windows, linux and mac. Anyway, I actually got a running > simulation > (just physical stuff) of a bunch of 4 wheeled robots just like what > we're modeling. ODE is fast and stable and it will allow us to do > some > great stuff. Right now I'm just working on the final piece of the > puzzle, the Evolve to ODE interface and meanwhile I'm cleaning up a > bunch of other stuff. I created a new project called Evolve3. I am > confident this will become our mainline codebase within a week or > two, > I've already made significant progress. > > This is the feature set I expect as a part of phase 1: > > 1. Mostly continuous world. No more rounds!!! > 2. Realistic physics. The bugs have 4 wheels (each motor controls > two wheels, kinda like a tank). > 3. Movement impedance will be replaced by the concept of friction > (I'm not 100% sure about this one, more of a try it out and see if > it makes sense type feature. Ridiculously simple to implement > however.) > > This is planned feature set for phase 2, subject to change > depending how > phase 1 works out: > > 1. Have a concept of "Sensor Space" where sensors are represented as > physical objects that are located somewhere on the bug chassis. > However instead of belonging to the same collision space that the > bugs belong to, they will have their own sensor space. Stimuli > objects (like aroma clouds) will also exist in this "Sensor > Space", so all we look for is the collision between a sensor and > its stimuli after each step of the simulation (btw, "looking for > collisions" just involves asking ODE to find them for you =) ). > These kind of operations are actually cheap in ODE, since it > treats collisions in each space completely separately. The real > motivation here is not so much for the benefit of using ODE to do > our sensing for us, but more for letting ODE transform the sensors > physical position and orientation for us. > 2. Three Dimensions. Technically phase 1 will be in 3 dimensions but > the environment will still be 2D (no cliffs or peaks, just flat > area). I have some ideas on this might be done but i want to save > it for after we have something working. > 3. We can replace our bresenham line calculations with a ray that > goes from the bug to the light and collides with objects and cubes > of impedance representing environment cells. Then we just add up > all the impedances from all of the collisions. > > Phase 3 and Beyond > > 1. Create a whole new concept of actuators. How about giving the > bugs actual arms, and letting the bugs control joint rotations. > This will give us a mechanism to allow bugs to grab on to each > other. We are only limited here by our own imagination. > > > Okay here's some of the bad news about this change, and we can > discuss > how important or unimportant these issues are: > > 1. The world won't wrap anymore. > 2. Audio stimuli needs to be rethought, since it won't quite work as > well anymore. I'm thinking to just represent audio and a sphere > which radius grows, and we can use the "Sensor Space" of phase 2 > to determine when a bug senses audio. We can use the ray method > to calculate audio impedances > > Tell me what you guys think =) > Rob > > > _______________________________________________ > Evolveproject-design mailing list > Evo...@li... > https://lists.sourceforge.net/lists/listinfo/evolveproject-design > > > __________________________________________________ > Do You Yahoo!? > Tired of spam? Yahoo! Mail has the best spam protection around > http://mail.yahoo.com > > ------------------------------------------------------------------------ > > _______________________________________________ > Evolveproject-design mailing list > Evo...@li... > https://lists.sourceforge.net/lists/listinfo/evolveproject-design > |
