  |
AIBO at Various Scales[2006.07.16] |
|
|
This report continues the previous LeggedSim integration in USARSim 2.x - Scale issue. I made more tests with the AIBO at 250uu/m, 80uu/m and 52.5uu/m scale and with QRIO at 80uu/m scale.
AIBO@52A.mpg
Penetration is too high to allow any evaluation.
AIBO@52B.mpg
AIBO does not penetrate anymore, however we can still see some slow movements. Probably this can be adjusted by writing custom inertia tensors.
AIBO@80.mpg
Here we can see the AIBO in the 250uu/m soccer map and in the physics test arena. At this scale AIBO performs slightly better. However I cannot further improve its behavior only by "blindly" adjusting Karma params, now I need to test it with the real AIBO code to see what to change. We'll make tests at both 52.5 and 80 scales.
AIBO@250.mpg
I was curious to see how the AIBO behavior would be changed by the KActorGravScale correction. As the scale is 250uu/m, the GravScale is 250/79.7 = 3.14. In the movie you can see that the AIBO is rather heavy (compared to the previous tests), even if the gravity force seems to be correct. The overall simulation, in any case, appears to be realistic.
QRIO@80.mpg
As my QRIO controller is very simple I cannot test it properly. However, at 80uu/m scale, it seems to behave correctly and its standing position is quite stable.
I think that the AIBO behavior can be tuned to work both at 52.5 and 80 scale, but not easily. It's more difficult to adjust Karma parameters at small scales, and the resulting simulation appears to be less precise (and requires more time? see Engine Load at Different Scales).
All these testing sessions are VERY time consuming and I have some doubt about their validity and utility. For example, even the 79.7uu/m scale factor changes slightly if you make the gravity test on different machines (it probably depends on the euler integrator that is less precise at larger time steps). I think that Karma could never be "perfectly" tuned.
I'm also anxious to proceed with the LeggedSim integration because it's taking too much time. Now I'll make the final tests with the AIBO (using real code this time) and I'll write another report when finished. After that, I hope that we'll agree on the scale factor in a short time. I know that I was not able to find the "best" scale to use, I only found that larger scales work better. To find an optimal scale we should automate the experiments and write scripts that could measure karma precision (how?) and visualization speed (FPSLog). However, as evaluating a physics engine is really difficult (and slightly PC dependent), I believe that we'll never know what is best. Besides, we already planned to update USARSim to UT2007 release, so it doesn't make sense to spend so much energies to find an answer that will certainly change.
Despite all, I still believe that a bigger scale could bring more precision in the simulation and help in map building. Perhaps 250uu/m is too much, but a value between 100 and 200 (150?) could be a good compromise (consider that there are robots that are even smaller than AIBO). If only someone else could test its robots at different scales we would have a better chance to guess it right. In any case, if you think that 52.5uu/m is the way to go, I'll conform myself to it and I'll try to do my best to tune AIBO accordingly.