Implementation:Google deepmind Dm control MJCF Observable

Knowledge Sources	Google_deepmind_Dm_control
Domains	Composer, Observables, MJCF
Last Updated	2026-02-15 04:00 GMT

Overview

The mjcf observable module defines observables that reference MJCF elements directly rather than by string name, enabling type-safe observation definitions that survive model composition and namespace changes.

Description

MJCFFeature takes an mjcf.Element (or iterable of elements) and a kind string (such as 'qpos', 'xpos'), using physics.bind() to access the corresponding attribute at runtime. It supports indexing via __getitem__ to slice specific components from the bound attribute. The _callable method returns a closure that, when called, reads the current value from the physics simulation.

MJCFCamera takes a <camera> MJCF element and renders images with configurable height, width, depth, segmentation, scene options, and render flag overrides. It provides a BoundedArray spec via the array_spec property with appropriate bounds and shapes for RGB (uint8, 3-channel), depth (float32, 1-channel), or segmentation (int32, 2-channel) modes. The _callable method returns a function that calls physics.render() with the configured parameters.

Both classes extend base.Observable and accept standard observable parameters including update_interval, buffer_size, delay, aggregator, and corruptor for integration with the Composer observation buffering system.

Usage

Use MJCFFeature to define observables for any physics quantity associated with MJCF elements (joint positions, body positions, sensor values). Use MJCFCamera to define pixel-based observables from cameras in the scene. These are the preferred observable types in Composer as they remain valid through model attachment and namespace prefixing.

Code Reference

Source Location

• Repository: Google_deepmind_Dm_control
• File: dm_control/composer/observation/observable/mjcf.py
• Lines: 1-276

Signature

class MJCFFeature(base.Observable):
    def __init__(self, kind, mjcf_element, update_interval=1,
                 buffer_size=None, delay=None,
                 aggregator=None, corruptor=None, index=None):
    def _callable(self, physics):
    def __getitem__(self, key):

class MJCFCamera(base.Observable):
    def __init__(self, mjcf_element, height=240, width=320,
                 update_interval=1, buffer_size=None, delay=None,
                 aggregator=None, corruptor=None, depth=False,
                 segmentation=False, scene_option=None,
                 render_flag_overrides=None):
    @property
    def height(self):
    @property
    def width(self):
    @property
    def depth(self):
    @property
    def segmentation(self):
    @property
    def scene_option(self):
    @property
    def render_flag_overrides(self):
    @property
    def array_spec(self):
    def _callable(self, physics):

Import

from dm_control.composer.observation.observable.mjcf import MJCFFeature
from dm_control.composer.observation.observable.mjcf import MJCFCamera

I/O Contract

Inputs (MJCFFeature)

Name	Type	Required	Description
kind	str	Yes	Name of a bound mjcf.Physics attribute (e.g., 'qpos', 'xpos', 'sensordata')
mjcf_element	mjcf.Element or iterable	Yes	The MJCF element(s) to observe
update_interval	int	No	Simulation steps between updates (default: 1)
buffer_size	int	No	Maximum buffer size (default: DEFAULT_BUFFER_SIZE)
delay	int	No	Additional simulation step delay (default: DEFAULT_DELAY)
aggregator	str or callable	No	Reduction operation over the buffer dimension
corruptor	callable	No	Function to corrupt/noise the observation
index	index	No	Index to apply to the attribute array

Inputs (MJCFCamera)

Name	Type	Required	Description
mjcf_element	mjcf.Element	Yes	A <camera> MJCF element
height	int	No	Rendered image height in pixels (default: 240)
width	int	No	Rendered image width in pixels (default: 320)
depth	bool	No	Render a depth image instead of RGB (default: False)
segmentation	bool	No	Render a segmentation mask (default: False)
scene_option	MjvOption	No	Custom visualization options
render_flag_overrides	dict	No	Rendering flag overrides

Outputs

Name	Type	Description
MJCFFeature observation	numpy.ndarray	The physics attribute values for the bound element(s)
MJCFCamera observation	numpy.ndarray	Rendered pixel array of shape (height, width, channels)
MJCFCamera.array_spec	specs.BoundedArray	Specification with shape, dtype, and bounds for the camera output

Usage Examples

from dm_control.composer.observation.observable.mjcf import MJCFFeature, MJCFCamera

# Observe joint positions for a set of MJCF joint elements
joint_obs = MJCFFeature('qpos', entity.mjcf_model.find_all('joint'))

# Observe only the first 3 components of a body's position
body_pos = MJCFFeature('xpos', entity.root_body)[0:3]

# Create a camera observable with depth rendering
camera_element = entity.mjcf_model.find('camera', 'front_camera')
camera_obs = MJCFCamera(
    mjcf_element=camera_element,
    height=480,
    width=640,
    depth=True,
)

# Create a segmentation camera
seg_camera = MJCFCamera(
    mjcf_element=camera_element,
    height=240,
    width=320,
    segmentation=True,
)