Implementation:Google deepmind Dm control Suite Humanoid CMU

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

Overview

Concrete tool for controlling a high-dimensional CMU humanoid model to stand, walk, or run provided by the dm_control Control Suite.

Description

The Humanoid CMU domain uses a more detailed humanoid model based on the Carnegie Mellon University motion capture skeleton. Compared to the standard Humanoid domain, this model has many more degrees of freedom and uses the thorax as the reference body instead of the torso. The Physics subclass provides methods for reading thorax uprightness (y-axis to world z-axis projection), head height, center-of-mass position and velocity, torso vertical orientation (from the thorax frame), joint angles (excluding the 7 root free-joint DOFs), and extremity positions (using left/right hand/foot naming with short prefixes 'l'/'r').

Three tasks are registered (all non-benchmarking): stand, walk, and run. All use the HumanoidCMU task class parameterized by move_speed. The stand task sets move_speed=0 and rewards standing upright with the head above 1.4 m while penalizing horizontal movement and large controls. The walk and run tasks set move_speed to 1 and 10 m/s respectively, rewarding horizontal locomotion at or above the target speed.

Episode initialization randomizes all limited and rotational joints to a collision-free configuration. All tasks use a default time limit of 20 seconds and a control timestep of 0.02 seconds. The reward structure mirrors the standard Humanoid domain but references thorax instead of torso for uprightness calculations.

Usage

Use this implementation for high-dimensional locomotion tasks with the CMU humanoid skeleton. Load via suite.load(domain_name='humanoid_CMU', task_name='stand') or the walk/run variants.

Code Reference

Source Location

• Repository: Google_deepmind_Dm_control
• File: dm_control/suite/humanoid_CMU.py
• Lines: 1-186

Signature

# Task factory functions
def stand(time_limit=20, random=None, environment_kwargs=None)
def walk(time_limit=20, random=None, environment_kwargs=None)
def run(time_limit=20, random=None, environment_kwargs=None)

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

# Task class
class HumanoidCMU(base.Task):
    def __init__(self, move_speed, 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_CMU', task_name='stand')

I/O Contract

Inputs

Name	Type	Required	Description
time_limit	float	No	Maximum episode duration in seconds (default 20).
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

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 thorax frame.
torso_vertical	numpy array (3,)	Z-projection of thorax orientation matrix.
com_velocity	numpy array (3,)	Center-of-mass velocity from sensor.
velocity	numpy array	Joint velocities.

Usage Examples

from dm_control import suite

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

# 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 walk and run tasks
env_walk = suite.load(domain_name='humanoid_CMU', task_name='walk')
env_run = suite.load(domain_name='humanoid_CMU', task_name='run')

Related Pages

• Principle:Google_deepmind_Dm_control_Control_Suite_Environment_Loading