Using Drake from Python

Background

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.

Drake is not tested regularly with Anaconda, so if you are using Anaconda you may experience compatibility hiccups; when asking for help, be sure to mention that Conda is involved.

Installation

Refer to Installation via Pip for how to install Drake’s releases using pip, or the general Installation instructions for alternative options.

Using the Python Bindings

Check Installation

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

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

If you are using Gurobi, you must either have it installed in the suggested location under /opt/... mentioned in Gurobi 10.0, 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 and the Robotic Manipulation 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 pydrake.systems.analysis import Simulator
from pydrake.systems.framework import DiagramBuilder

builder = DiagramBuilder()
plant, _ = AddMultibodyPlantSceneGraph(builder, 0.0)
Parser(plant).AddModels(
    FindResourceOrThrow("drake/examples/pendulum/Pendulum.urdf"))
plant.Finalize()
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)
Parser(plant).AddModels(
    FindResourceOrThrow("drake/examples/pendulum/Pendulum.urdf"))
plant.Finalize()
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 = pydrake.systems.framework.DiagramBuilder()
plant, _ = pydrake.multibody.plant.AddMultibodyPlantSceneGraph(builder, 0.0)
pydrake.multibody.parsing.Parser(plant).AddModels(
  pydrake.common.FindResourceOrThrow(
      "drake/examples/pendulum/Pendulum.urdf"))
plant.Finalize()
diagram = builder.Build()
simulator = pydrake.systems.analysis.Simulator(diagram)

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 pydrake.systems.primitives import Adder, Adder_
>>> print(Adder)
<class 'pydrake.systems.primitives.Adder_𝓣float𝓤'>
>>> print(Adder_)
<TemplateClass pydrake.systems.primitives.Adder_>
>>> from pydrake.autodiffutils import AutoDiffXd
>>> from pydrake.symbolic import Expression
>>> print(Adder_[float])
<class 'pydrake.systems.primitives.Adder_𝓣float𝓤'>
>>> print(Adder_[AutoDiffXd])
<class 'pydrake.systems.primitives.Adder_𝓣AutoDiffXd𝓤'>
>>> print(Adder_[Expression])
<class 'pydrake.systems.primitives.Adder_𝓣Expression𝓤'>

In debugging output like the class repr shown above, you might encounter the unicode letters 𝓣 and 𝓤. These are used for “name mangling” of template types; we need to use “name mangling” to obey Python’s class and function naming rules. If you see a mangled name, you can read it using the following legend: a 𝓣 stands for an open bracket ([), a 𝓤 stands for a close bracket (]), a 𝓬 stands for a comma (,), and a 𝓹 stands for a dot (.).

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)
<pydrake.systems.primitives.Adder_𝓣float𝓤 object at 0x...>
>>> print(adder.ToAutoDiffXd())
<pydrake.systems.primitives.Adder_𝓣AutoDiffXd𝓤 object at 0x...>
>>> print(adder.ToSymbolic())
<pydrake.systems.primitives.Adder_𝓣Expression𝓤 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:

DiagramBuilder<T>::AddSystem<SystemType>(args...)
MultibodyPlant<T>::AddJoint<JointType>(args...)
MultibodyPlant<T>::AddFrame<FrameType>(args...)

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:

InputPort<T>::Eval<ValueType>(context)
GeometryProperties::AddProperty<ValueType>(group_name, name, value)

will become the following in Python:

InputPort_[T].Eval(context)
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])
        sys.stdout.flush()
        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)

    @functools.wraps(func)
    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.
@traced
def main():
    insert_awesome_code_here()


if __name__ == "__main__":
    reexecute_if_unbuffered()
    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.

Additionally, you can also decorate your function to break on an exception so you can get a REPL (which works in a terminal or in a Jupyter notebook) to actively inspect the context (like dbstop if error in MATLAB).

from contextlib import contextmanager
import pdb
import sys
import traceback


@contextmanager
def launch_pdb_on_exception():
    """
    Provides a context that will launch interactive pdb console automatically
    if an exception is raised.

    Example usage with @iex decorator shorthand below:

        @iex
        def my_bad_function():
            x = 1
            assert False

        my_bad_function()
        # Should bring up debugger at `assert` statement.
    """
    # Adapted from:
    # https://github.com/gotcha/ipdb/blob/fc83b4f5f/ipdb/__main__.py#L219-L232

    try:
        yield
    except Exception:
        traceback.print_exc()
        _, _, tb = sys.exc_info()
        pdb.post_mortem(tb)
        # Resume original execution.
        raise


# Mirror `@ipdb.iex` decorator. See docs for `launch_pdb_on_exception()`.
iex = launch_pdb_on_exception()

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.