Implementation:Google deepmind Dm control Suite Humanoid

Metadata	Value
Implementation	Suite Humanoid
Domain	Reinforcement_Learning, Control
Source	Google_deepmind_Dm_control
Last Updated	2026-02-15 04:00 GMT

Overview

Concrete tool for controlling a 3D humanoid to stand, walk, or run provided by the dm_control Control Suite.

Description

The Humanoid domain models a full 3D humanoid body with 21 degrees of freedom and multiple actuated joints. The Physics subclass provides methods for querying the torso uprightness (z-axis projection), head height, center-of-mass position and velocity, torso vertical orientation, joint angles (excluding the 7 root DOFs of the free joint), and end effector (hands and feet) positions in the egocentric torso frame.

Four tasks are registered: three benchmarking tasks (stand, walk, run) and one extra task (run_pure_state). All use the Humanoid task class parameterized by move_speed and pure_state. The stand task sets move_speed=0 and rewards the agent for keeping the head above 1.4 m, maintaining torso uprightness, using small controls, and not moving horizontally. The walk and run tasks set move_speed to 1 and 10 m/s respectively, rewarding horizontal locomotion at or above the target speed. The run_pure_state variant provides raw MuJoCo positions and velocities as observations instead of egocentric features.

Episode initialization randomizes all limited and rotational joints to a collision-free configuration. All tasks use a default time limit of 25 seconds and a control timestep of 0.025 seconds.

Usage

Use this implementation for high-dimensional 3D locomotion benchmarks. Load via suite.load(domain_name='humanoid', task_name='stand') or the walk/run variants.

Code Reference

Source Location

• Repository: Google_deepmind_Dm_control
• File: dm_control/suite/humanoid.py
• Lines: 1-207

Signature

# Task factory functions
def stand(time_limit=25, random=None, environment_kwargs=None)
def walk(time_limit=25, random=None, environment_kwargs=None)
def run(time_limit=25, random=None, environment_kwargs=None)
def run_pure_state(time_limit=25, random=None, environment_kwargs=None)

# Physics subclass
class Physics(mujoco.Physics):
    def torso_upright(self)                # z-axis projection of torso
    def head_height(self)                  # height of the head
    def center_of_mass_position(self)      # CoM position
    def center_of_mass_velocity(self)      # CoM velocity from sensor
    def torso_vertical_orientation(self)   # z-projection of torso orientation
    def joint_angles(self)                 # joint positions sans root DOFs
    def extremities(self)                  # hand/foot positions in torso frame

# Task class
class Humanoid(base.Task):
    def __init__(self, move_speed, pure_state, random=None)
    def initialize_episode(self, physics)
    def get_observation(self, physics)
    def get_reward(self, physics)

Import

from dm_control import suite

env = suite.load(domain_name='humanoid', task_name='walk')

I/O Contract

Inputs

Name	Type	Required	Description
time_limit	float	No	Maximum episode duration in seconds (default 25).
random	int, numpy.random.RandomState, or None	No	Random seed or RNG instance for reproducibility.
environment_kwargs	dict or None	No	Additional keyword arguments forwarded to the Environment constructor.

Outputs

Name	Type	Description
environment	dm_control.rl.control.Environment	A fully initialised environment conforming to the dm_env.Environment interface.

Observations (egocentric mode)

Key	Type	Description
joint_angles	numpy array	Joint positions excluding the 7 root free-joint DOFs.
head_height	float	Height of the head above the ground.
extremities	numpy array (12,)	Left/right hand and foot positions in torso frame.
torso_vertical	numpy array (3,)	Z-projection of torso orientation matrix.
com_velocity	numpy array (3,)	Center-of-mass velocity from sensor.
velocity	numpy array	Joint velocities.

Observations (pure state mode)

Key	Type	Description
position	numpy array	Full generalized positions.
velocity	numpy array	Full generalized velocities.

Usage Examples

from dm_control import suite

# Load the walk task
env = suite.load(domain_name='humanoid', task_name='walk')

# Run an episode
time_step = env.reset()
while not time_step.last():
    action = env.action_spec().generate_value()
    time_step = env.step(action)

# Load the run task
env_run = suite.load(domain_name='humanoid', task_name='run')

# Load with pure state observations
env_pure = suite.load(domain_name='humanoid', task_name='run_pure_state')

Related Pages

• Principle:Google_deepmind_Dm_control_Control_Suite_Environment_Loading