Build Pinocchio with HPP-FCL support option

[sholybest]

First, thank you to everyone reading this message.
When I ran a meshcat display code that shows a model from example-robot-data, I got some errors same with those in this link #1371, "No to_python (by-value) converter found for C++ type ...".

As the answer suggested at the link, I compiled pinocciho adding hpp-fcl support option as following.
cmake .. -DBUILD_WITH_COLLISION_SUPPORT=ON -DBUILD_WITH_OPENMP_SUPPORT=ON -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS="-march=native"

However, when BUILD_WITH_COLLISION_SUPPORT option was added, I got an error message as in the attached image below.

The versions of the dependencies are as follows. When I installed them with pip install or robotpkg, these same versions didn't make problems and displayed the robot well. How can I solve this error??
Eigen: 3.3.7
eigenpy: 3.4.0
Pinocchio: 2.7.0
hpp-fcl: 2.4.4
Boost: 1.83.0

Thank you.

[jorisv]

Hello @sholybest,

Has I remember, this is an issue between boost serialization and Eigen.
It has been fixed in this hpp-fcl PR.

Unfortunately, this patch is not released. We will release the new hpp-fcl version soon, but in the mean time, you can build hpp-fcl from source on the following commit : 40a636b40e6ca3caf27b3d580de0697044c406af

[sholybest]

I wanted to enhance the performance of the whole project, by compiling all the dependencies with appropriate options for my native architecture, rather than getting pre-compiled packages.

I expected cmake to automatically find packages in /usr/local directory..
After getting your advice, I set CMAKE_PREFIX_PATH environment variable, and added redundantly in CmakeList.txt the followings:
include_directories(/usr/local/include)
link_directories(/usr/local/lib)
link_libraries(hpp-fcl)

Then cmake found /usr/local/libhpp-fcl.so as in the image below.

But still same error remains..
When I pip installed hpp-fcl (at /home/nam/.local/lib/python3.8/site-packages/cmeel.prefix/), the error was removed and pinocchio could be built with hpp-fcl support option. The robot model could be successfully displayed in meshcat.
It did not make a big difference in computation time, so I would be satisfied with this option.

Maybe my system is messed up during many trials.. Sorry for bothering you with this unskilled question, and I appreciate your consideration and patience. I'll looking forward the new releases!

[nim65s]

You probably need to isolate your build process from anything else. If you have multiple setups configured at the same time, or some setup in a standard location (like /usr or /usr/local), those will prevent you from correctly configuring, compiling and running your projects.

Maybe you can try to do that in a Dockerfile, to ensure isolation and reproducibility ? Otherwise, can you share your full clone / configure / build process with a script, so we can review it ?

[sholybest]

Sorry for being late and that I don't have any experience with Dockerfiles.. but I can leave the procedure I did.

CPU: Intel i5-6600k
OS: Ubuntu20.04

1. Boost downloaded from https://www.boost.org/

tar xf boost_1_83_0.tar.bz2
cd boost_1_83_0

./bootstrap.sh --prefix=/usr/local --with-python=python3
sudo ./b2 -j2 toolset=gcc cxxflags="-O3 -march=native" threading=multi link=shared install 

2. urdfdom-headers

pip install cmeel-urdfdom-headers

3. eigenpy
   Source code of 3.4.0 downloaded from https://github.com/stack-of-tasks/eigenpy/releases

cd eigenpy 
mkdir build && cd build
cmake .. -DBUILD_WITH_ACCELERATE_SUPPORT=ON -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS="-march=native 
-fopenmp" 
sudo make -j2 
sudo make install

4. hpp-fcl

git clone --recursive https://github.com/humanoid-path-planner/hpp-fcl
cd hpp-fcl
git reset --hard 40a636b40e6ca3caf27b3d580de0697044c406af
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS="-march=native -fopenmp" 
sudo make -j2 
sudo make install

5. pinocchio
   Added this in CmakeLists.txt

include_directories(/usr/local/include)
link_directories(/usr/local/lib)
link_libraries(hpp-fcl)

git clone --recursive https://github.com/stack-of-tasks/pinocchio.git
cd pinocchio
mkdir build && cd build
cmake .. -DBUILD_WITH_COLLISION_SUPPORT=ON -DBUILD_WITH_OPENMP_SUPPORT=ON -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS="-march=native" 
sudo make -j2
sudo make install

[nim65s]

Thanks, this help.

First, I would like to ask what is the rationale for downloading boost instead of using eq. libboost-all-dev ubuntu package and not taking the latest release ? and cmeel-urdfdom-headers instead of liburdfdom-headers-dev ubuntu package ? And not taking the latest eigenpy release ?

Also, I would say that if you are seeking better performances, I think upgrading your CPU is a way simpler option than trying to build everything with march=native by yourself. 6600K is nice, but it is almost 10 years old.

[sholybest]

Thank you very much for your concerns and supports.
I wanted to do the real-time MPC simulation of quadruped with Box-FDDP solver of Crocoddyl library. When I pip installed Crocoddyl, the same versions mentioned above were installed automatically. I just followed them and there's no other reasons for choosing the versions. I didn't know about urdfdom-headers at all, but I could get it by pip install. That's all..

The SBC implemented in our quadruped robot platform is very old one, and its performance is similar to that of my desktop. This spec was sufficient for our previous MPC controller, with Single Rigid Body model. So I wanted to see even the most simple locomotion code of Crocoddyl would not be affordable. I would try my best and discuss with my team members and advisor, to upgrade the CPU. Do you think this spec is not enough for full-body DDP of quadruped?