RussTedrake
commented
9 years ago Just to be clear on how you got here: NonlinearProgram is trying to auto-differentiate your constraint function for you (hence the TaylorVar). This is happening because of the call in your stack to geval (short for gradient eval) which tries autodiff by default if you have not supplied the gradients to the requested order manually. This is a powerful feature. :)
You've hit a case in forwardKin where we were properly supporting the autodiff. I just sent a pull-request with a one line fix. Unfortunately, it means that you found some code that is not covered by our autodiff unit tests, so there may be a few more one-line fixes like this to make it work. Try it and let me know.
Other options are to set the gradient method for the constraint to be 'numerical' or to implement the 'user' gradients yourself. You might look at the help in geval to understand better.
On Jul 8, 2015, at 11:51 PM, aespielberg notifications@github.com wrote:
Hi,
Sorry for the dumb question, but, when I try to add function handle constraints, my state is converted to a TaylorVar. For example, I make a constraint:
constraint = FunctionHandleConstraint(lb, ub, 10, @(state)positionDiff(plant, frame_name, point_in_frame, target_point, state), 0); (I don't think all the constraint parameterization here is important, but if you want I can post the function.)
Then, I add it to my program:
prog = prog.addConstraint(constraint, [last_x_inds; prog.param_inds(:)]); However, in the function positionDiff (which I have defined), state gets passed in as a TaylorVar.
I'm curious what the correct way to use this TaylorVar, then. I had planned on manipulating the state simply as a double (I thought I had done this correctly in the past but I never used addConstraint, only addStateConstraint, so I'm unsure if it's different). I can't pass in TaylorVar to plant.setParams(), so for my parameterization purposes, I can't use it like this. I also keep getting the following warning when I try to treat it like a double:
Warning: converting taylorvar to double even though it has non-zero gradients. gradient information will be lost!
In TaylorVar/double (line 47) In RigidBodyManipulator/forwardKin (line 127) In positionDiff (line 43) In constructPositionDiffConstraint>@(state)positionDiff(plant,frame_name,point_in_frame,target_point,state) (line 3) In FunctionHandleConstraint/constraintEval (line 44) In Constraint>@(varargin)obj.constraintEval(varargin{:}) (line 123) In geval (line 171) In Constraint/eval (line 123) In NonlinearProgram/objectiveAndNonlinearConstraints (line 581) In NonlinearProgram/snopt/snopt_userfun (line 1314) In snoptUserfun (line 6) In NonlinearProgram/snopt (line 1377) In NonlinearProgram/solve (line 973) In DirectTrajectoryOptimization/solveTraj (line 188) In KinematicDirtranTest/solveTraj (line 82) In AcrobotPlantTest/swingUpTrajectoryKinematic (line 200) In runSwingUpKinematicTest (line 6) Is there a correct usage here? I imagine the gradient information is useful in some way for returning df in my constraint.
— Reply to this email directly or view it on GitHub.
aespielberg
Hi,
Sorry for the dumb question, but, when I try to add function handle constraints, my state is converted to a TaylorVar. For example, I make a constraint:
(I don't think all the constraint parameterization here is important, but if you want I can post the function.)
Then, I add it to my program:
However, in the function positionDiff (which I have defined), state gets passed in as a TaylorVar.
I'm curious what the correct way to use this TaylorVar, then. I had planned on manipulating the state simply as a double (I thought I had done this correctly in the past but I never used addConstraint, only addStateConstraint, so I'm unsure if it's different). I can't pass in TaylorVar to plant.setParams(), so for my parameterization purposes, I can't use it like this. I also keep getting the following warning when I try to treat it like a double:
Is there a correct usage here? I imagine the gradient information is useful in some way for returning df in my constraint. I should make an explicit note that the warning is in using TaylorVar in forwardKin, which might not be a big deal since forwardKin returns a gradient anyway. I still don't know the proper way to use it with setParams though. I have semi-solved the problem by directly casting to double.