Geometric Algebra For RObotics in Python

This library provides a geometric algebra tools targeted towards robotics applications. It includes various computations for the kinematics and dynamics of serial manipulators as well as optimal control.

It is based on gafro, a C++ library relying on templates to efficiently implement the geometric algebra operations.

Note that only the Conformal Geometric Algebra part of gafro is available in pygafro.

Please visit https://gitlab.com/gafro in order to find the entire gafro software stack.

Installation using pip (pre-compiled binaries)

pip install pygafro

Wheels for Linux (x64_86 & arm64) and MacOS (arm64) are available (for Python 3.8 to 3.13).

For other platforms, pygafro is compiled from sources.

Should you want to have access to mesh and texture files for the robots, you can install the optional package pygafro-assets by running either one of those commands:

pip install pygafro-assets
pip install pygafro[assets]

Note that pygafro doesn't provide any rendering functions, it only indicates which mesh file to use for each link.

Installation using pip (compilation from sources)

Due to the template-based nature of gafro (see Differences between gafro and pygafro below), the compilation of pygafro can take a long time. Additionally, using clang instead of gcc is highly recommended: gcc requires more memory resources when compiling pygafro, which can become problematic on lower-end computers.

Using the default compiler of your computer

pip install pygafro

Forcing the usage of clang

(assuming that clang is installed at /usr/bin/clang)

export CC=/usr/bin/clang
export CXX=/usr/bin/clang++
pip install pygafro

Installation with ROS2

Add PyGafro in your colcon workspace and build it with:

CC=clang CXX=clang++ USE_COLCON=1 colcon build

Installation from source

(works either in a conda or virtual environment)

Requirements:

• numpy

Due to the template-based nature of gafro (see Differences between gafro and pygafro below), the compilation of pygafro can take a long time. Additionally, using clang instead of gcc is highly recommended: gcc requires more memory resources when compiling pygafro, which can become problematic on lower-end computers.

Using the default compiler of your computer

git clone
cd pygafro
mkdir build && cd build
cmake ..
make # or for example "make -j4" if you have enough resources
make install

Forcing the usage of clang

(assuming that clang is installed at /usr/bin/clang)

git clone
cd pygafro
mkdir build && cd build
cmake -DCMAKE_CXX_COMPILER=/usr/bin/clang++ -DCMAKE_C_COMPILER=/usr/bin/clang ..
make # or for example "make -j4" if you have enough resources
make install

Usage

Multivectors

from pygafro import Multivector
from pygafro import Point
from pygafro import Motor

# create a multivector that corresponds to a Euclidean vector
vector = Multivector.create(['e1', 'e2', 'e3'], [1.0, 2.0, 3.0])

# create a point (a specialized multivector subclass)
point = Point(1.0, 2.0, 3.0)

# create a random motor
motor = Motor.Random()

# apply the motor to our multivectors
vector2 = motor.apply(vector)
point2 = motor.apply(point)

# geometric product
result = vector * point

# inner product
result = vector | point

# outer product
result = vector ^ point

Robots

from pygafro import FrankaEmikaRobot

panda = FrankaEmikaRobot()

position = panda.getRandomConfiguration()

# forward kinematics: compute the motor at the end-effector
ee_motor = panda.getEEMotor(position)

Differences between gafro and pygafro

gafro being based on C++ templates, only the classes and operations you are effectively using are compiled into your software.

This versatility cannot be achieved in a Python library: we cannot instantiate the templates at runtime, nor can we realistically instantiate all the possible combinations at compile time.

A compromise was choosen: a subset of multivectors (using sensible blades combinations) are instantiated and compiled, and other blades combinations are supported through a Python class that internally use a C++ multivector with more blades and transparently use a mask to only expose the blades requested by the user.

Thus, creating a multivector is done using the following helper function:

# using values
vector = Multivector.create(['e1', 'e2', 'e3'], [1.0, 2.0, 3.0])

# using only the list of blades
vector = Multivector.create(['e1', 'e2', 'e3', 'ei', 'e123i'])

Background

You can find the accompanying article here and more information on our website.

How to cite

If you use gafro in your research, please cite:

@article{loewGeometricAlgebraOptimal2023,
  title = {Geometric {{Algebra}} for {{Optimal Control}} with {{Applications}} in {{Manipulation Tasks}}},
  author = {L\"ow, Tobias and Calinon, Sylvain},
  date = {2023},
  journal = {IEEE Transactions on Robotics},
  doi = {10.1109/TRO.2023.3277282}
}

Acknowledgments

The development of pygafro is actively supported by the Idiap Research Institute.