Trying to replicate the Biped Gap Crossing scenario about towr HOT 3 CLOSED

Good question. Exactly, this is also related to the way rviz shapes are being used to approximate the actual height map used by the optimizer. If you look at

PS: You can also model the gap directly as the rviz picture, with non-continuous heights. This can work as well, however having the gradient information of the parabola available helps the solver converge quicker.

from towr.

Hi @ferrolho, yes, this looks good, i think the optimizer is doing exactly as it's supposed to and only the visualization in rviz is off.

The height map the optimizer user internally, with a width of 1m is encoded here:

The class that tries to replicate the internal height map with rviz shapes is faultily using only 0.5m and is coded here:

So for now just make sure these coincide and I'm actually working on how not to duplicate that information, but at the same time introduce no dependencies on rviz in the core solver class. I'll also fix this bug in that PR and update this issue then.

from towr.

Hey @awinkler, thank you for the prompt reply!

This was indeed an issue and after setting double l_gap = 1.0; result looks better:

However, I am still a bit confused... This is the log of the optimiser I get on my slow laptop:

Number of Iterations....: 69

                                   (scaled)                 (unscaled)
Objective...............:   0.0000000000000000e+00    0.0000000000000000e+00
Dual infeasibility......:   7.9408063850289635e-06    7.9408063850289635e-06
Constraint violation....:   5.2894000180652583e-12    1.2967404927621828e-11
Complementarity.........:   1.8847416957485208e-06    1.8847416957485208e-06
Overall NLP error.......:   7.9408063850289635e-06    7.9408063850289635e-06


Number of objective function evaluations             = 225
Number of objective gradient evaluations             = 64
Number of equality constraint evaluations            = 225
Number of inequality constraint evaluations          = 225
Number of equality constraint Jacobian evaluations   = 71
Number of inequality constraint Jacobian evaluations = 71
Number of Lagrangian Hessian evaluations             = 0
Total CPU secs in IPOPT (w/o function evaluations)   =      1.221
Total CPU secs in NLP function evaluations           =     25.192

EXIT: Optimal Solution Found.

It says "Optimal Solution Found" and does not report any constraint to have been violated. However, in the animation above we can see that, at least visually, this does not seem to be the case:

from towr.