Python Bindings

Introduction

Starting with version 2.1.2, MuJoCo comes with native Python bindings that are developed in C++ using pybind11. Unlike previous Python bindings, these are officially supported by the MuJoCo development team and will be kept up-to-date with the latest developments in MuJoCo itself.

The Python bindings are distributed as the mujoco package on PyPI. These are low-level bindings that are meant to give as close to a direct access to the MuJoCo library as possible. However, in order to provide an API and semantics that developers would expect in a typical Python library, the bindings deliberately diverge from the raw MuJoCo API in a number of places, which are documented throughout this page.

DeepMind’s dm_control reinforcement learning library (which prior to version 1.0.0 implemented its own MuJoCo bindings based on ctypes) has been updated to depend on the mujoco package and continues to be supported by DeepMind. Changes in dm_control should be largely transparent to users of previous versions, however code that depended directly on its low-level API may need to be updated. Consult the migration guide for detail.

For mujoco-py users, we include notes below to aid migration.

Installation

The package can be installed from PyPI via

pip install mujoco

A copy of the MuJoCo library is provided as part of the package and does not need to be downloaded or installed separately.

Building from source

Source code for the Python bindings are available in the python top-level directory in MuJoCo’s GitHub repository. Developers wishing to build the bindings from source should work with a full clone of the Git repository, run the make_sdist.sh script to generate a source distribution (sdist) tarball, then run pip wheel name_of_sdist.tar.gz to build the libraries and generate a wheel. The make_sdist.sh script generates additional C++ header files that are needed to build the bindings, and also pulls in other required files from elsewhere in the repository outside the python directory into the sdist.

CMake and a C++17 compiler are needed to build the bindings from source.

Basic usage

Once installed, the package can be imported via import mujoco. Structs, functions, constants, and enums are available directly from the top-level mujoco module.

Structs

MuJoCo data structures are exposed as Python classes. In order to conform to PEP 8 naming guidelines, struct names begin with a capital letter, for example mjData becomes mujoco.MjData in Python.

All structs other than mjModel have constructors in Python. For structs that have an mj_defaultFoo-style initialization function, the Python constructor calls the default initializer automatically, so for example mujoco.MjOption() creates a new mjOption instance that is pre-initialized with mj_defaultOption. Otherwise, the Python constructor zero-initializes the underlying C struct.

Structs with a mj_makeFoo-style initialization function have corresponding constructor overloads in Python, for example mujoco.MjvScene(model, maxgeom=10) in Python creates a new mjvScene instance that is initialized with mjv_makeScene(model, [the new mjvScene instance], 10) in C. When this form of initialization is used, the corresponding deallocation function mj_freeFoo/mj_deleteFoo is automatically called when the Python object is deleted. The user does not need to manually free resources.

The mujoco.MjModel class does not a have Python constructor. Instead, we provide three static factory functions that create a new mjModel instance: mujoco.MjModel.from_xml_string, mujoco.MjModel.from_xml_path, and mujoco.MjModel.from_binary_path. The first function accepts a model XML as a string, while the latter two functions accept the path to either an XML or MJB model file. All three functions optionally accept a Python dictionary which is converted into a MuJoCo Virtual file system for use during model compilation.

Functions

MuJoCo functions are exposed as Python functions of the same name. Unlike with structs, we do not attempt to make the function names PEP 8-compliant, as MuJoCo uses both underscores and CamelCases. In most cases, function arguments appear exactly as they do in C, and keyword arguments are supported with the same names as declared in mujoco.h. Python bindings to C functions that accept array input arguments expect NumPy arrays or iterable objects that are convertible to NumPy arrays (e.g. lists). Output arguments (i.e. array arguments that MuJoCo expect to write values back to the caller) must always be writeable NumPy arrays.

In the C API, functions that take dynamically-sized arrays as inputs expect a pointer argument to the array along with an integer argument that specifies the array’s size. In Python, the size arguments are omitted since we can automatically (and indeed, more safely) deduce it from the NumPy array. When calling these functions, pass all arguments other than array sizes in the same order as they appear in mujoco.h, or use keyword arguments. For example, mj_jac should be called as mujoco.mj_jac(m, d, jacp, jacr, point, body) in Python.

The bindings releases the Python Global Interpreter Lock (GIL) before calling the underlying MuJoCo function. This allows for some thread-based parallelism, however users should bear in mind that the GIL is only released for the duration of the MuJoCo C function itself, and not during the execution of any other Python code.

Note

One place where the bindings do offer added functionality is the top-level mj_step function. Since it is often called in a loop, we have added an additional nstep argument, indicating how many times the underlying mj_step should be called. If not specified, nstep takes the default value of 1. The following two code snippets perform the same computation, but the first one does so without acquiring the GIL in between subsequent physics steps:

mj_step(model, data, nstep=20)
for _ in range(20):
  mj_step(model, data)

Enums and constants

MuJoCo enums are available as mujoco.mjtEnumType.ENUM_VALUE, for example mujoco.mjtObj.mjOBJ_SITE. MuJoCo constants are available with the same name directly under the mujoco module, for example mujoco.mjVISSTRING.

Minimal example

import mujoco

XML=r"""
<mujoco>
  <asset>
    <mesh file="gizmo.stl"/>
  </asset>
  <worldbody>
    <body>
      <freejoint/>
      <geom type="mesh" name="gizmo" mesh="gizmo"/>
    </body>
  </worldbody>
</mujoco>
"""

ASSETS=dict()
with open('/path/to/gizmo.stl', 'rb') as f:
  ASSETS['gizmo.stl'] = f.read()

model = mujoco.MjModel.from_xml_string(XML, ASSETS)
data = mujoco.MjData(model)
while data.time < 1:
  mujoco.mj_step(model, data)
  print(data.geom_xpos)

Named access

Most well-designed MuJoCo models assign names to objects (joints, geoms, bodies, etc.) of interest. When the model is compiled down to an mjModel instance, these names become associated with numeric IDs that are used to index into the various array members. For convenience and code readability, the Python bindings provide “named access” API on MjModel and MjData. Each name_fooadr field in the mjModel struct defines a name category foo.

For each name category foo, mujoco.MjModel and mujoco.MjData objects provide a method foo that takes a single string argument, and returns an accessor object for all arrays corresponding to the entity foo of the given name. The accessor object contains attributes whose names correspond to the fields of either mujoco.MjModel or mujoco.MjData but with the part before the underscore removed. For example:

  • m.geom('gizmo') returns an accessor for arrays in the MjModel object m associated with the geom named “gizmo”.

  • m.geom('gizmo').rgba is a NumPy array view of length 4 that specifies the RGBA color for the geom. Specifically, it corresponds to the portion of m.geom_rgba[4*i:4*i+4] where i = mujoco.mj_name2id(m, mujoco.mjtObj.mjOBJ_GEOM, 'gizmo').

Additionally, the Python API define a number of aliases for some name categories corresponding to the XML element name in the MJCF schema that defines an entity of that category. For example, m.joint('foo') is the same as m.jnt('foo'). A complete list of these aliases are provided below.

The accessor for joints is somewhat different that of the other categories. Some mjModel and mjData fields (those of size size nq or nv) are associated with degrees of freedom (DoFs) rather than joints. This is because different types of joints have different numbers of DoFs. We nevertheless associate these fields to their corresponding joints, for example through d.joint('foo').qpos and d.joint('foo').qvel, however the size of these arrays would differ between accessors depending on the joint’s type.

Named access is guaranteed to be O(1) in the number of entities in the model. In other words, the time it takes to access an entity by name does not grow with the number of names or entities in the model. (This is currently not the case for the mj_name2id function, which performs a linear scan.)

For completeness, we provide here a complete list of all name categories in MuJoCo, along with their corresponding aliases defined in the Python API.

  • body

  • jnt or joint

  • geom

  • site

  • cam or camera

  • light

  • mesh

  • skin

  • hfield

  • tex or texture

  • mat or material

  • pair

  • exclude

  • eq or equality

  • tendon or ten

  • actuator

  • sensor

  • numeric

  • text

  • tuple

  • key or keyframe

Rendering

MuJoCo itself expects users to set up a working OpenGL context before calling any of its mjr_ rendering routine. The Python bindings provide a basic class mujoco.GLContext that helps users set up such a context for offscreen rendering. To create a context, call ctx = mujoco.GLContext(max_width, max_height). Once the context is created, it must be made current before MuJoCo rendering functions can be called, which you can do so via ctx.make_current(). Note that a context can only be made current on one thread at any given time, and all subsequent rendering calls must be made on the same thread.

The context is freed automatically when the ctx object is deleted, but in some multi-threaded scenario it may be necessary to explicitly free the underlying OpenGL context. To do so, call ctx.free(), after which point it is the user’s responsibility to ensure that no further rendering calls are made on the context.

Once the context is created, users can follow MuJoCo’s standard rendering, for example as documented in the Visualization section.

Error handling

MuJoCo reports irrecoverable errors via the mju_error mechanism, which immediately terminates the entire process. Users are permitted to install a custom error handler via the mju_user_error callback, but it too is expected to terminate the process, otherwise the behavior of MuJoCo after the callback returns is undefined. In actuality, it is sufficient to ensure that error callbacks do not return to MuJoCo, but it is permitted to use longjmp to skip MuJoCo’s call stack back to the external callsite.

The Python bindings utilises longjmp to allow it to convert irrecoverable MuJoCo errors into Python exceptions of type mujoco.FatalError that can be caught and processed in the usual Pythonic way. Furthermore, it installs its error callback in a thread-local manner using a currently private API, thus allowing for concurrent calls into MuJoCo from multiple threads.

Callbacks

MuJoCo allows users to install custom callback functions to modify certain parts of its computation pipeline. For example, mjcb_sensor can be used to implement custom sensors, and mjcb_control can be used to implement custom actuators. Callbacks are exposed through the function pointers prefixed mjcb_ in mujoco.h.

For each callback mjcb_foo, users can set it to a Python callable via mujoco.set_mjcb_foo(some_callable). To reset it, call mujoco.set_mjcb_foo(None). To retrieve the currently installed callback, call mujoco.get_mjcb_foo(). (The getter should not be used if the callback is not installed via the Python bindings.) The bindings automatically acquire the GIL each time the callback is entered, and release it before reentering MuJoCo. This is likely to incur a severe performance impact as callbacks are triggered several times throughout MuJoCo’s computation pipeline and is unlikely to be suitable for “production” use case. However, it is expected that this feature will be useful for prototyping complex models.

Alternatively, if a callback is implemented in a native dynamic library, users can use ctypes to obtain a Python handle to the C function pointer and pass it to mujoco.set_mjcb_foo. The bindings will then retrieve the underlying function pointer and assign it directly to the raw callback pointer, and the GIL will not be acquired each time the callback is entered.

Migration Notes for mujoco-py

In mujoco-py, the main entry point is the MjSim class. Users constuct a stateful MjSim instance from an MJCF model (similar to dm_control.Physics), and this instance holds references to an mjModel instance and its associated mjData. In contrast, the MuJoCo Python bindings (mujoco) take a more low-level approach, as explained above: following the design principle of the C library, the mujoco module itself is stateless, and merely wraps the underlying native structs and functions.

While a complete survey of mujoco-py is beyond the scope of this document, we offer below implementation notes for a non-exhaustive list of specific mujoco-py features:

mujoco_py.load_model_from_xml(bstring)

This factory function constructs a stateful MjSim instance. When using mujoco, the user should call the factory function mujoco.MjModel.from_xml_* as described above. The user is then responsible for holding the resulting MjModel struct instance and explicitly generating the corresponding MjData by calling mujoco.MjData(model).

sim.reset(), sim.forward(), sim.step()

Here as above, mujoco users needs to call the underlying library functions, passing instances of MjModel and MjData: mujoco.mj_resetData(model, data), mujoco.mj_forward(model, data), and mujoco.mj_step(model, data).

sim.get_state(), sim.set_state(state), sim.get_flattened_state(), sim.set_state_from_flattened(state)

The MuJoCo library’s computation is deterministic given a specific input, as explained in the Programming section. mujoco-py implements methods for getting and setting some of the relevant fields (and similarly dm_control.Physics offers methods that correspond to the flattened case). mujoco do not offer such abstraction, and the user is expected to get/set the values of the relevant fields explicitly.

sim.model.get_joint_qvel_addr(joint_name)

This is a convenience method in mujoco-py that returns a list of contiguous indices corresponding to this joint. The list starts from model.jnt_qposadr[joint_index], and its length depends on the joint type. mujoco doesn’t offer this functionality, but this list can be easily constructed using model.jnt_qposadr[joint_index] and xrange.

sim.model.*_name2id(name)

mujoco-py creates dicts in MjSim that allow for efficient lookup of indices for objects of different types: site_name2id, body_name2id etc. These functions replace the function mujoco.mj_name2id(model, type_enum, name) whose current implementation is inefficient. mujoco offers a different approach for using entity names – named access, as well as access to the native mj_name2id.

sim.save(fstream, format_name)

This is the one context in which the MuJoCo library (and therefore also mujoco) is stateful: it holds a copy in memory of the last XML that was compiled, which is used in mujoco.mj_saveLastXML(fname). Note that mujoco-py’s implementation has a convenient extra feature, whereby the pose (as determined by sim.data’s state) is transformed to a keyframe that’s added to the model before saving. This extra feature is not currently available in mujoco.

Code Sample: open-loop rollout

We include a code sample showing how to add additional C/C++ functionality, exposed as a Python module via pybind11. The sample, implemented in rollout.cc and wrapped in rollout.py, implements a common use case where tight loops implemented outside of Python are beneficial: rolling out a trajectory (i.e., calling mj_step() in a loop), given an intial state and sequence of controls, and returning subsequent states and sensor values. The canonical usage form is

state, sensordata = rollout.rollout(model, data, initial_state, ctrl)

initial_state is a nstate x nqva array, with nstate initial states of length nqva, where nqva = model.nq + model.nv + model.na is the size of the full MuJoCo mechanical state: positions (data.qpos), velocities (data.qvel) and actuator activations (data.act). ctrl is a nstate x nstep x nu array of control sequences.

The rollout function is designed to be completely stateless, so all inputs of the stepping pipeline are set and any values already present in the given MjData instance will have no effect on the output. In order to facilitate this, all inputs including time and qacc_warmstart are set to default values, as are auxillary controls (qfrc_applied, xfrc_applied and mocap_{pos,quat}). These can also be optionally set by the user.

Since the Global Interpreter Lock can be released, this function can be efficiently threaded using Python threads. See the test_threading function in rollout_test.py for an example of threaded operation.