Tracikpy

This repo contains Python bindings for TracIK based on the ROS packages and SWIG bindings provided by TRACLabs here (see paper for implementation details) and the updates provided by Clemens Eppner here. For now, it only supports Ubuntu operating systems. The goal of this repo is to encapsulate these bindings into a Python package that can be easily installed using pip (with numpy as the only Python dependency) and can be used independently of ROS. This package does still contain several system-level dependencies (eigen, orocos-kdl, nlopt, urdfdom, and kdl_parser); if you have already installed ROS then these will be installed already. If not, you can install using apt-get:

sudo apt-get install libeigen3-dev liborocos-kdl-dev libkdl-parser-dev liburdfdom-dev libnlopt-dev

For Ubuntu 20.04, libnlopt-cxx-dev is also needed to install the C++ bindings for nlopt.

Main differences from original library/bindings:

ROS and Boost dependencies removed (replaced with C++ standard library calls).
Installable via pip, taking advantage of SWIG support in setuptools.Extension.
Added forward kinematics.
Simplified Python wrapper.
Added basic unit tests through pytest

Install

Clone the repo and install using pip:

git clone https://github.com/mjd3/tracikpy.git
pip install tracikpy/

That's it!

Quickstart

Here is an example script showing how this package can be used to calculate inverse kinematics:

import numpy as np
from tracikpy import TracIKSolver

ee_pose = np.array([[ 0.0525767 , -0.64690764, -0.7607537 , 0.        ],
                    [-0.90099786, -0.35923817,  0.24320937, 0.2       ],
                    [-0.43062577,  0.67265031, -0.60174996, 0.4       ],
                    [ 0.        ,  0.        ,  0.        , 1.        ]])

ik_solver = TracIKSolver(
    "data/franka_panda.urdf",
    "panda_link0",
    "panda_hand",
)
qout = ik_solver.ik(ee_pose, qinit=np.zeros(ik_solver.number_of_joints))
print(qout)

which yields:

[-0.05331206 -1.75956216  0.6304369  -3.00941705  1.28849325  0.46002026  2.58126884]

Note that the qinit argument is optional; if you do not include it then it will be set to a configuration chosen uniformly at random between the robot's joint limits.

You can also check the solution using forward kinematics:

ee_out = ik_solver.fk(qout)
ee_diff = np.linalg.inv(ee_pose) @ ee_out
trans_err = np.linalg.norm(ee_diff[:3, 3], ord=1)
angle_err = np.arccos(np.trace(ee_diff[:3, :3] - 1) / 2)
assert trans_err < 1e-3
assert angle_err < 1e-3 or angle_err - np.pi < 1e-3

which should not output any assertion errors since the pose is close to the desired pose.

Unit Tests

To run the unit tests for this project, install the repo with the test option and run pytest:

cd /path/to/tracikpy
pip install -e .[test]
pytest

TODO

GitHub Actions CI for MacOS and Windows
Migrate SWIG std_vector.i templates to numpy.i (see here) or migrate to pybind11.
Integrate CIBuildWheel?

Acknowledgments

Many thanks to TRACLabs (Patrick Beeson and Barrett Ames) for the initial TracIK library, to Sammy Pfeiffer for the initial Python bindings, and to Clemens Eppner for initial cleanup to remove ROS dependencies.