pyb_utils: utilities for PyBullet

This is a collection of utilities I've found useful for working with PyBullet, including:

• Collision detection: conveniently set up shortest distance computations and collision checking between arbitrary objects in arbitrary configurations with PyBullet. See the accompanying blog post.
• Ghost objects: add purely visual objects to the simulation, optionally attached to another body.
• Camera: virtual camera from which to get RGBA, depth, segmentation, and point cloud data. Also provides video recording using OpenCV.
• Convenience class for easily creating rigid bodies.
• Versions of some PyBullet functions that return named tuples, for easy field access.
• Basic quaternion functions.

Install and run

This package supports Python 3.8–3.11. It has been tested on Ubuntu 16.04, 18.04, and 20.04.

From pip

pip install pyb_utils

From source

Clone the repository:

git clone https://github.com/adamheins/pyb_utils
cd pyb_utils

Install using poetry:

poetry install
poetry run python examples/collision_detection_example.py  # for example

Or using pip:

python -m pip install .

Documentation

The project's documentation is available here.

Usage and examples

This package provides a few basic quality-of-life utilities. First, PyBullet represents rotations using quaternions (in [x, y, z, w] order). We provide a few helper routines to create quaternions about the principal axes, convert quaternions to rotation matrices, and to rotate points (using scipy under the hood):

>>> import pyb_utils
>>> q = pyb_utils.quatz(np.pi / 2) # 90 deg rotation about z-axis
>>> q
array([0., 0., 0.70710678, 0.70710678])

>>> pyb_utils.quaternion_to_matrix(q)  # convert to rotation matrix
array([[-0.,  -1.,  0.],
       [ 1.,  -0.,  0.],
       [ 0.,   0.,  1.]])

>>> pyb_utils.quaternion_multiply(q, q)  # rotate two quaternions together
array([0, 0, 1, 0])                      # 180 deg rotate about z

>>> pyb_utils.quaternion_rotate(q, [1, 0, 0])  # rotate a point
array([0, 1, 0])

Second, we provide a simple class to quickly create rigid bodies programmatically, which is useful for adding basic objects to manipulate or act as obstacles:

>>> import pybullet as pyb
>>> import pyb_utils

>>> pyb.connect(pyb.GUI)

# create a 1x1x1 cube at the origin
>>> box = pyb_utils.BulletBody.box(position=[0, 0, 0], half_extents=[0.5, 0.5, 0.5])

# put a ball on top
>>> ball = pyb_utils.BulletBody.sphere(position=[0, 0, 1.5], radius=0.5)

# now put it somewhere else
>>> ball.set_pose(position=[2, 0, 0.5])

Third, we wrap some PyBullet functions to return named tuples, rather than normal tuples. When the tuples have 10+ fields in them, it is rather helpful to have names! The names and arguments of these functions are the same as the underlying PyBullet ones, to make swapping effortless. Continuing our previous example:

# built-in PyBullet method
# the output is not easy to read!
>>> pyb.getDynamicsInfo(box.uid, -1)
(1.0,
 0.5,
 (0.16666666666666666, 0.16666666666666666, 0.16666666666666666),
 (0.0, 0.0, 0.0),
 (0.0, 0.0, 0.0, 1.0),
 0.0,
 0.0,
 0.0,
 -1.0,
 -1.0,
 2,
 0.001)

# switch to the pyb_utils version
# now we can access fields by name
>>> info = pyb_utils.getDynamicsInfo(box.uid, -1)
>>> info.mass
1.0
>>> info.localInertiaDiagonal
(0.16666666666666666, 0.16666666666666666, 0.16666666666666666),

The functions we've wrapped in this way are getClosestPoints, getConstraintInfo, getContactPoints, getDynamicsInfo, getJointInfo, getJointState(s), and getLinkState(s). There are two differences from the vanilla PyBullet API. The first is that in pyb_utils getJointInfo also accepts an optional argument decode, which will convert the byte strings returned by PyBullet to the specified encoding. For example, decode="utf8". The second difference is that in pyb_utils getLinkState(s) will always return LinkState tuples with 8 fields, even if computeLinkVelocity=False. When computeLinkVelocity=False, then worldLinkLinearVelocity and worldLinkAngularVelocity are both set to None.

And there's more! You can find example scripts of all of this package's utilities in the examples/ directory:

• rigid bodies
• camera
• collision detection
• ghost objects
• named tuples
• video

Video Output

Writing a video with the VideoRecorder defaults to using the mp4v codec, which is widely supported but (at least on my computer running Ubuntu 20.04) does not play natively in web browsers. The availability of codecs depends on what is compiled into the version of OpenCV you have installed (i.e., the one backing the cv2 Python module); using an alternative codec may require a different version of OpenCV.

Known issues

Feel free to open issues (or better yet, a pull request!) if you find a problem. Currently known issues:

• Ghost objects sometimes flicker (spooky, but undesirable). This is probably because they are updated by directly changing the object pose; we cannot have them updated automatically by, e.g., constraints since they are not dynamic objects (and we wouldn't want them to be; then they would influence the simulation).
• The field name localInerialPos in the DynamicsInfo named tuple is spelled incorrectly. This will be fixed in a future major version.
• The deprecated GhostSphere class will be removed in a future major version. Use GhostObject.sphere instead.

Development

• Run tox to run the tests.
• Sphinx is used to build the documentation. With Sphinx installed, run make html in the docs directory.

License

MIT