Using Drake from Python

A substantial subset of the Drake C++ functionality is available from Python. The Drake Python bindings are generated using pybind11, which means that every function or class which is exposed to C++ has been explicitly enumerated in one of the source files inside the bindings/pydrake folder. These bindings are installed as a single package called pydrake.


The Python environment supplied by Anaconda is neither tested nor supported for building and using the Drake Python bindings.


Before attempting installation, please review the supported configurations to know what versions of Python are supported for your platform.

Binary Installation for Python

First, download and extract an available binary package.

As an example, here is how to download and extract one of the latest releases to /opt (where <platform> could be bionic, focal, or mac):

curl -o drake.tar.gz<platform>.tar.gz
rm -rf /opt/drake
tar -xvzf drake.tar.gz -C /opt

Ensure that you have the system dependencies:


Next, ensure that your PYTHONPATH is properly configured.

Ubuntu 18.04 (Bionic):

export PYTHONPATH=/opt/drake/lib/python3.6/site-packages:${PYTHONPATH}

Ubuntu 20.04 (Focal):

export PYTHONPATH=/opt/drake/lib/python3.8/site-packages:${PYTHONPATH}


export PYTHONPATH=/opt/drake/lib/python3.8/site-packages:${PYTHONPATH}

See below for usage instructions.

Inside virtualenv

At present, Drake is not installable via pip. However, you can still incorporate its install tree into a virtualenv FHS-like environment.

An example, where you should replace <venv_path> and <platform>:

# Setup drake, and run prerequisites.
curl -o drake.tar.gz<platform>.tar.gz
mkdir -p <venv_path>
tar -xvzf drake.tar.gz -C <venv_path> --strip-components=1
# - You may need `sudo` here.

# Setup a virtualenv over the drake install.
python3 -m virtualenv -p python3 <venv_path> --system-site-packages


You can extract Drake into an existing virtualenv tree if you have already run install_prereqs; however, you should ensure that you have run install_prereqs. Before you do this, you should capture / freeze your current requirements to reproduce your environment if there are conflicts.

To check if this worked, follow the instructions as shown below, but either:

  • Use <venv_path>/bin/python instead of python3, or
  • Source <venv_path>/bin/activate in your current shell session.

Building the Python Bindings

To use the Python bindings from Drake externally, we recommend using CMake. As an example:

git clone
mkdir drake-build
cd drake-build
cmake ../drake
make -j

Please note the additional CMake options which affect the Python bindings:

  • -DWITH_GUROBI={ON, [OFF]} - Build with Gurobi enabled.
  • -DWITH_MOSEK={ON, [OFF]} - Build with MOSEK enabled.
  • -DWITH_SNOPT={ON, [OFF]} - Build with SNOPT enabled.

{...} means a list of options, and the option surrounded by [...] is the default option. An example of building pydrake with both Gurobi and MOSEK, without building tests:


You will also need to have your PYTHONPATH configured correctly.

Ubuntu 18.04 (Bionic):

cd drake-build
export PYTHONPATH=${PWD}/install/lib/python3.6/site-packages:${PYTHONPATH}

Ubuntu 20.04 (Focal):

cd drake-build
export PYTHONPATH=${PWD}/install/lib/python3.8/site-packages:${PYTHONPATH}


cd drake-build
export PYTHONPATH=${PWD}/install/lib/python3.8/site-packages:${PYTHONPATH}

Using the Python Bindings

Check Installation

After following the above install steps, check to ensure you can import pydrake.

python3 -c 'import pydrake; print(pydrake.__file__)'


If you are using Gurobi, you must either have it installed in the suggested location under /opt/... mentioned in Gurobi 9.0.2, or you must ensure that you define the ${GUROBI_HOME} environment variable, or specify ${GUROBI_INCLUDE_DIR} via CMake.

What’s Available from Python

You should first browse the Python API to see what modules are available. The most up-to-date high-level demonstrations of what can be done using pydrake are in Drake’s Tutorials and the Underactuated Robotics Textbook.

You can also see lower-level usages of the API in the pydrake unit tests themselves, which you can find inside of the drake/bindings/python/pydrake/**/test folders in the Drake source code.

Here’s an example snippet of code from pydrake:

from pydrake.common import FindResourceOrThrow
from pydrake.multibody.parsing import Parser
from pydrake.multibody.plant import AddMultibodyPlantSceneGraph
from import Simulator
from import DiagramBuilder

builder = DiagramBuilder()
plant, _ = AddMultibodyPlantSceneGraph(builder, 0.0)
diagram = builder.Build()
simulator = Simulator(diagram)

If you are prototyping code in a REPL environment (such as IPython / Jupyter) and to reduce the number of import statements, consider using pydrake.all to import a subset of symbols from a flattened namespace or import all modules automatically. If you are writing non-prototype code, avoid using pydrake.all; for more details, see help(pydrake.all).

In all cases, try to avoid using from pydrake.all import *, as it may introduce symbol collisions that are difficult to debug.

The above example, but using pydrake.all:

from pydrake.all import (
    AddMultibodyPlantSceneGraph, DiagramBuilder, FindResourceOrThrow,
    Parser, Simulator)

builder = DiagramBuilder()
plant, _ = AddMultibodyPlantSceneGraph(builder, 0.0)
diagram = builder.Build()
simulator = Simulator(diagram)

An alternative is to use pydrake.all to import all modules, but then explicitly refer to each symbol:

import pydrake.all

builder =
plant, _ = pydrake.multibody.plant.AddMultibodyPlantSceneGraph(builder, 0.0)
diagram = builder.Build()
simulator =

Differences with C++ API

In general, the Python API should be close to the C++ API. There are some exceptions:

C++ Class Template Instantiations in Python

When you define a general class template, e.g. template <typename T> class Value, something like Value<std::string> is called the instantiation.

For certain C++ templated types, they are exposed in Pythons also as templates; the parameter types (in this case, T) are the Python-equivalent types to the C++ type. Some examples:

C++ Python
std::string str
double float, np.double, np.float64, ctypes.c_double
drake::AutoDiffXd pydrake.autodiffutils.AutoDiffXd
drake::symbolic::Expression pydrake.symbolic.Expression

Thus, the instantiation Value<std::string> will be bound in Python as Value[str].

Scalar Types

Most classes in the Systems framework and in the multibody dynamics computational framework are templated on a scalar type, T. For convenience (and backwards compatibility) in Python, a slightly different binding convention is used.

For example, Adder<T> is a Systems primitive which has a user-defined number of inputs and outputs a single port which is the sum of all of the inputs.

In C++, you would access the instantiations using Adder<double>, Adder<AutoDiffXd>, and Adder<Expression> for common scalar types.

In Python, Adder actually refers to the “default” instantiation, the Adder<double> C++ class. To access other instantiations, you should add an _ to the end of the C++ class name to get the Python template and then provide the parameters in square braces, [...]. In this example, you should use Adder_[T].

To illustrate, you can print out the string representations of Adder, Adder_, and some of its instantiations in Python:

>>> from import Adder, Adder_
>>> print(Adder)
<class '[float]'>
>>> print(Adder_)
>>> from pydrake.autodiffutils import AutoDiffXd
>>> from pydrake.symbolic import Expression
>>> print(Adder_[float])
<class '[float]'>
>>> print(Adder_[AutoDiffXd])
<class '[AutoDiffXd]'>
>>> print(Adder_[Expression])
<class '[Expression]'>

Additionally, you may convert an instance (if the conversion is available) using System_[T].ToAutoDiffXd and System_[T].ToSymbolic:

>>> adder = Adder(num_inputs=1, size=1)
>>> print(adder)
<[float] object at 0x...>
>>> print(adder.ToAutoDiffXd())
<[AutoDiffXd] object at 0x...>
>>> print(adder.ToSymbolic())
<[Expression] object at 0x...>

C++ Function and Method Template Instantiations in Python

The above section indicates that C++ types are generally registered with Python, and a similar approach could be used for function and method templates. However, these templates usually fit a certain pattern and can be Pythonized in such a way that simplifies implementation, but may change the “feel” of the signature.

Two common (non-metaprogramming) applications of templated functions and methods present in Drake are emplace-like functionality (using parameter packs) and type erasure. However, Python doesn’t literally support these C++ language features. So, in binding them, they get “Pythonized”.

C++ APIs which use parameter packs, such as:


will become the following in Python:

DiagramBuilder_[T].AddSystem(SystemType(args, ...))
MultibodyPlant_[T].AddJoint(JointType(args, ...))
MultibodyPlant_[T].AddFrame(FrameType(args, ...))

where the *Type tokens are replaced with the concrete type in question (e.g. Adder_[T], RevoluteJoint_[T], FixedOffsetFrame_[T]).

Similarly, type-erasure C++ APIs that look like:

GeometryProperties::AddProperty<ValueType>(group_name, name, value)

will become the following in Python:

GeometryProperties.AddProperty(group_name, name, value)

Debugging with the Python Bindings

You may encounter issues with the Python Bindings that may arise from the underlying C++ code, and it may not always be obvious what the root cause is.

The first step to debugging is to consider running your code using the trace module. It is best practice to always have a main() function, and have a if __name__ == "__main__" clause. If you do this, then it is easy to trace. As an example:

def reexecute_if_unbuffered():
    """Ensures that output is immediately flushed (e.g. for segfaults).
    ONLY use this at your entrypoint. Otherwise, you may have code be
    re-executed that will clutter your console."""
    import os
    import shlex
    import sys
    if os.environ.get("PYTHONUNBUFFERED") in (None, ""):
        os.environ["PYTHONUNBUFFERED"] = "1"
        argv = list(sys.argv)
        if argv[0] != sys.executable:
            argv.insert(0, sys.executable)
        cmd = " ".join([shlex.quote(arg) for arg in argv])
        os.execv(argv[0], argv)

def traced(func, ignoredirs=None):
    """Decorates func such that its execution is traced, but filters out any
     Python code outside of the system prefix."""
    import functools
    import sys
    import trace
    if ignoredirs is None:
        ignoredirs = ["/usr", sys.prefix]
    tracer = trace.Trace(trace=1, count=0, ignoredirs=ignoredirs)

    def wrapped(*args, **kwargs):
        return tracer.runfunc(func, *args, **kwargs)

    return wrapped

# NOTE: You don't have to trace all of your code. If you can identify a
# single function, then you can just decorate it with this. If you're
# decorating a class method, then be sure to declare these functions above
# it.
def main():

if __name__ == "__main__":


If you are developing in Drake and are using the drake_py_unittest macro, you can specify the argument --trace=user to get the same behavior.

This generally should help you trace where the code is dying. However, if you still need to dig in, you can build the bindings in debug mode, without symbol stripping, so you can debug with gdb or lldb:

cmake -DCMAKE_BUILD_TYPE=Debug ../drake


If you have SNOPT enabled (either -DWITH_SNOPT=ON or -DWITH_ROBOTLOCOMOTION_SNOPT=ON), symbols will still be stripped.

For Developers

If you are developing Python bindings, please see the Doxygen page Python Bindings which provides information on programming conventions, documentation, tips for debugging, and other advice.