Change to Double Precision? #100
Comments
|
First data point, memory. A four segment tetraspine (18 strings) requires: Runtime is unchanged, both simulate 10 seconds of time at 500 Hz in ~5.2 real seconds (without graphics). |
|
I re-ran the ICRA 2015 paper trials on the MuscleNP branch and observed the following:
Given (2), I favor moving to double precision as the simulator's default configuration. I'll leave this issue open for a little while longer just in case others have tests they can think of or comments. |
|
Closing based on previous post |
Merging in MuscleNP, new muscle constructors and examples. This closes #96 Note the following about this merge: 1) The simulator is now running at double precision by default. We recommend migrating, as it increases simulator accuracy for a minor increase in memory usage (see: #100). In order to migrate you need to do one of the following: a) delete your env and build directories and re-run setup.sh, followed by a clean build (bin/build.sh) b) Rebuild bullet at double precision (cd to the bullet directory: env/build/bullet, run "ccmake .", turn double precision on, generate new cmake files, run make install), then do the same thing to NTRT (ccmake build, toggle option on and generate, run bin/build.sh) 2) The order of the parameters in tgBaseString::config has changed. The third parameter is now pretension, in units of force, to allow pretension in cables without a controller. Rotation has been moved to the end, history is fourth, and everything else has been left unchanged. If your code was in master, I've updated it for you. If you were just specifying stiffness and damping you don't need to do anything, but otherwise check tgBaseString::config and make sure everything is as you expect. There's a new example of the 3 bar prism (3_prism_no_control) showing how to use this new pretension parameter. Finally - there's a new cable model in dev - MuscleNP, which includes contact dynamics. There's several examples of its use in my dev directory (btietz), the most interesting of which is TCContact, which demonstrates their use on hilly terrain. Changing your models to use contact dynamics is just a matter of using dev/tgMultiPointStringInfo instead of tgLinearStringInfo, so it should be easy to play with. Its still in beta, so let me know if you find bugs (or post them to github).
Prior to merging in #99, we need to determine if double precision breaks any of the validation results we have previously published. MuscleNP functions better with it (more stable) and the results from 6WCSCM I have tested move slightly further. A comparison to the ICRA 2015 results, as well as other validation results as available is required before making the change.
The text was updated successfully, but these errors were encountered: