acrobot dircol example fails with --config snopt_fortran, but succeeds with the precompiled binaries

[russtedrake]

acrobot dircol example fails with --config snopt_fortran, but succeeds with the precompiled binaries

[RussTedrake]

i've confirmed that changing my bazel settings on the local install from --config=snopt_fortran back to --config=snopt fixes the issue on this machine. this confirms for me that it is a difference in snopt between versions.

[jwnimmer-tri]

I guess we're not using the fortran snopt in the precompiled binaries yet?

That is correct; RobotLocomotion/drake#10422 is the overall tracking issue.

... it is a difference in snopt between versions.

In your builds that failed/passed when toggling --config=snopt_fortran/--config=snopt, which source release of SNOPT were you testing the acrobot dircol against? Assuming you were using Bazel builds, the question is what was env SNOPT_PATH set to?

We're going to need to categorize this failure as either one of the "currently known bugs in SNOPT Fortran" shown in RobotLocomotion/drake#10422, or else file a new issue for this failure mode if its different. It's possible that @hongkai-dai has some insight in that regard, or might be able to suggest some tolerance or tuning options that can mitigate this change.

[RussTedrake]

i have been testing on CI against the "latest" "continuous" drake-binaries
https://github.com/RussTedrake/underactuated/blob/master/drake_version.json

I also run nightly tests against "nightly" (I pick early AM to be after the drake nightlies come out). have not seen the failure. My CI does not set SNOPT_PATH. https://github.com/RussTedrake/underactuated/blob/master/.circleci/config.yml

On my machine (where I was seeing failures when I flipped to snopt_fortran), I build directly from master, and do set SNOPT_PATH="git".

[jwnimmer-tri]

... testing on CI against the "latest" "continuous" drake-binaries; also ... against "nightly"; have not seen the failure. My CI does not set SNOPT_PATH.

With precompiled Drake, you are getting SNOPT 7.4 source archive built using F2C mode.

On my machine (where I was seeing failures when I flipped to snopt_fortran), I build directly from master, and do set SNOPT_PATH="git".

That's the SNOPT 7.4 source archive (as of RobotLocomotion/drake#10430) built using Fortran mode.

Given that data, it sounds to me like this issue is neither RobotLocomotion/drake#10423 (unbounded) nor RobotLocomotion/drake#10424 (bug in 7.6 fortran sources, not 7.4 fortran sources).

[RussTedrake]

To be clear, I don't think we can declare this new instance as a bug. The nonlinear optimization is not guaranteed to find a solution to this problem. But it is a degradation in performance. 7.4 seems to solve it well, and 7.6 does not.

[jwnimmer-tri]

... and 7.6 does not.

nit: The above data only shows that 7.4/Fortran does not solve it well (but 7.4/F2C does). We haven't tested 7.6/Fortran yet.

[RussTedrake]

I believe this has been resolved, as since RobotLocomotion/drake#10422 has been completed. will reopen if more testing results in a different understanding.