Implementation:Google deepmind Dm control Variation Distributions

Knowledge Sources	Google_deepmind_Dm_control
Domains	Reinforcement_Learning, Domain_Randomization
Last Updated	2026-02-15 04:00 GMT

Overview

Standard statistical distribution classes conforming to the Variation API, providing the primary source of stochasticity for domain randomization in dm_control Composer environments.

Description

The distributions module implements a hierarchy of statistical distribution classes built on the Variation abstract base class. The central abstraction is the Distribution class, which provides a common framework for parametrized distributions. Subclasses implement a _callable(random_state) method that returns the appropriate NumPy random sampling function. When called, Distribution.__call__ evaluates its constructor arguments (which may themselves be Variation objects), determines the sample shape from initial_value (unless single_sample=True), and invokes the sampling function.

The module provides ten concrete distribution classes: Uniform (continuous uniform), UniformInteger (discrete uniform), UniformChoice (uniform selection from a list), Normal (Gaussian), LogNormal, Exponential, Poisson, and Bernoulli (binary coin flip via binomial). Additionally, UniformPointOnSphere samples uniformly distributed points on the unit 3D sphere by normalizing Gaussian random vectors, and BiasedRandomWalk implements a discrete-time Ornstein-Uhlenbeck process that produces temporally correlated noise with configurable standard deviation and correlation timescale.

All distribution parameters can themselves be Variation objects, enabling hierarchical randomization where distribution parameters are sampled from other distributions. The single_sample flag controls whether the distribution draws one scalar sample or an array matching the shape of the initial value.

Usage

Use these distributions to randomize MJCF model attributes, physics parameters, positions, orientations, and other properties in Composer tasks. Bind them to element attributes via MJCFVariator or PhysicsVariator, or evaluate them directly via variation.evaluate.

Code Reference

Source Location

Repository: Google_deepmind_Dm_control
File: dm_control/composer/variation/distributions.py
Lines: 1-259

Signature

class Distribution(base.Variation, metaclass=abc.ABCMeta):
    def __init__(self, *args, **kwargs):
        ...
    @abc.abstractmethod
    def _callable(self, random_state):
        ...

class Uniform(Distribution):
    def __init__(self, low=0.0, high=1.0, single_sample=False):
        ...

class UniformInteger(Distribution):
    def __init__(self, low, high=None, single_sample=False):
        ...

class UniformChoice(Distribution):
    def __init__(self, choices, single_sample=False):
        ...

class UniformPointOnSphere(base.Variation):
    def __init__(self, single_sample=False):
        ...

class Normal(Distribution):
    def __init__(self, loc=0.0, scale=1.0, single_sample=False):
        ...

class LogNormal(Distribution):
    def __init__(self, mean=0.0, sigma=1.0, single_sample=False):
        ...

class Exponential(Distribution):
    def __init__(self, scale=1.0, single_sample=False):
        ...

class Poisson(Distribution):
    def __init__(self, lam=1.0, single_sample=False):
        ...

class Bernoulli(Distribution):
    def __init__(self, prob=0.5, single_sample=False):
        ...

class BiasedRandomWalk(base.Variation):
    def __init__(self, stdev=0.1, timescale=10.):
        ...

Import

from dm_control.composer.variation import distributions
from dm_control.composer.variation.distributions import Uniform, Normal, BiasedRandomWalk

I/O Contract

Inputs (Distribution subclasses)

Name	Type	Required	Description
Distribution-specific params	Variation or numeric	Yes	Parameters of the distribution (e.g., `low`, `high` for Uniform)
single_sample	bool	No	If `True`, draw one scalar sample regardless of `initial_value` shape (default `False`)

Inputs (BiasedRandomWalk)

Name	Type	Required	Description
stdev	float	No	Standard deviation of the output sequence (default 0.1); must be >= 0
timescale	float	No	Number of timesteps for correlation decay (default 10.0); must be >= 0

Outputs

Name	Type	Description
return	scalar or `numpy.ndarray`	Sampled value(s); shape matches `initial_value` unless `single_sample=True`

Distribution Summary

Class	NumPy Function	Parameters	Description
`Uniform`	`random_state.uniform`	low, high	Continuous uniform distribution
`UniformInteger`	`random_state.randint`	low, high	Discrete uniform distribution
`UniformChoice`	`random_state.choice`	choices	Uniform selection from a list
`Normal`	`random_state.normal`	loc, scale	Gaussian distribution
`LogNormal`	`random_state.lognormal`	mean, sigma	Log-normal distribution
`Exponential`	`random_state.exponential`	scale	Exponential distribution
`Poisson`	`random_state.poisson`	lam	Poisson distribution
`Bernoulli`	`random_state.binomial(1, ...)`	prob	Bernoulli (coin flip) distribution
`UniformPointOnSphere`	normalized Gaussians	(none)	Uniform unit vector on the 3D sphere
`BiasedRandomWalk`	Ornstein-Uhlenbeck process	stdev, timescale	Temporally correlated noise

Usage Examples

from dm_control.composer.variation import distributions
import numpy as np

rng = np.random.RandomState(42)

# Sample from a uniform distribution
uniform = distributions.Uniform(low=-1.0, high=1.0)
value = uniform(initial_value=None, current_value=None, random_state=rng)

# Sample matching the shape of an initial value
initial = np.array([1.0, 2.0, 3.0])
values = uniform(initial_value=initial, current_value=None, random_state=rng)
# values has shape (3,)

# Sample a single scalar regardless of initial value shape
single = distributions.Uniform(low=0.0, high=1.0, single_sample=True)
scalar = single(initial_value=initial, current_value=None, random_state=rng)

# Sample a random point on the unit sphere
sphere = distributions.UniformPointOnSphere()
point = sphere(random_state=rng)  # 3D unit vector

# Create a biased random walk for correlated noise
walk = distributions.BiasedRandomWalk(stdev=0.1, timescale=10.)
noise_1 = walk(random_state=rng)
noise_2 = walk(random_state=rng)  # Correlated with noise_1

# Bernoulli distribution for binary randomization
coin = distributions.Bernoulli(prob=0.3)
flip = coin(random_state=rng)  # 0 or 1

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment