This is a library in C++ that implements Probabilistic Motion Primitives algorithms. References :
- Paraschos A, Daniel C, Peters J, Neumann G. Probabilistic movement primitives. Advances in neural information processing systems. 2013. [pdf]
- Paraschos A, Daniel C, Peters J, Neumann G. Using probabilistic movement primitives in robotics. Autonomous Robots. 2018 Mar;42(3):529-51. [pdf].
Initially, it was inspired in the following python implementations:
- C++ 11
- CMake 3.11
- Eigen 3.2.92
- Doxygen 1.8.11 (optional)
- Python (optional)
(Other versions of CMake, Eigen, Doxygen will probably work, but those are the tested ones so far)
git clone --recursive https://gitlab.inria.fr/H2020-AnDy/promp.git
cd promp
mkdir build
cd build
cmake ..
If you wish to compile the python library, set BUILD_PYTHON_MODULE to ON before compiling the library.
If you wish to build API documentation, set BUILD_DOCUMENTATIONS to ON before compiling the library (this requires doxygen and doxybook2).
Then, proceed to make and install:
make
sudo make install
If while executing your compiled targets you get an error like the following:
cannot open shared object file: No such file or directory
Please try running ldconfig before running again:
sudo ldconfig
At the moment, the library has been tested on Ubuntu 16.04 and 18.04.
You can refer to the source code of some examples in the examples folder to understand the usage of the library. Their binaries are automatically generated after the installation.
The api is described in doc/api.md
In the folder etc/demos/ you can find many recordings of human whole-body postures (i.csv i=1,2,...) and/or body segments poses (pi.csv i=1,2,...) for several motions. With these you can train and play with your own ProMPs. Have fun!
For an example usage in Java, see: src/test/java/us/ihmc/promp/test/SinglePrompExample.java. This example is a replica of the C++ example examples/single_promp.cpp.
If you make any changes to the C++ code, you should re-generate the Java bindings. Utilize the scripts provided in this project's root directory. These scripts will generate
the shared library files required at runtime (.so/.dll/.dylib) as well as the mapped Java code in src/main/generated-java.
To generate the Java bindings and shared libraries, for Linux, you should utilize Docker and the Docker build script in the root directory, generateJavaMappingsDocker.sh.
Using the Docker build script/the Docker image, you will ensure that the native libraries it produces are compatible with older distros (Ubuntu 20.04; glibc 2.31).
Before running the Docker build script, ensure your user can access Docker functions (normally, it's root-only access).
$ sudo groupadd docker
$ sudo usermod -aG docker $USER
$ newgrp docker
To generate the Java bindings and shared libraries for Windows, you should utilize the build script in the root directory, generateJavaMappings.bat.
This script assumes you have a few things installed:
- Visual Studio 2022
- CMake
- JDK 17
Python module is built using pybind11. Be sure to get it with git submodule update --init --recursive before using cmake.
Eigen matrices are converted to numpy matrices. Because of the different interal data storage order, the conversion between them implies a performance overhead.
After sudo make install, the module is automatically installed to Python site-lib folder.
In your python code, you can just import the module and use it as in C++. You can find a .py script in the example folder.
import promp
...
file_list = ["t1.csv", "t2.csv"]
dofs = [1,2,3]
trajectory_group = promp.TrajectoryGroup()
trajectory_group.load_csv_trajectories(file_list, dofs)
trajectory_group.normalize_length()
my_promp = promp.Promp(trajectory_group.trajectories(), 20, 0.05)
gen_traj = my_promp.generate_trajectory()
gen_stddev = my_promp.gen_traj_std_dev()
gen_covariance = my_promp.generate_trajectory_covariance()The development of this software is partially supported by the European Project H2020 An.Dy.