Implementation:Google deepmind Dm control Scaled Actuators

Knowledge Sources	Google_deepmind_Dm_control
Domains	Robotics, Physics Simulation
Last Updated	2026-02-15 04:00 GMT

Overview

The Scaled Actuators module provides utility functions for creating MuJoCo position and velocity actuators whose control signals are affinely mapped from a normalized range to desired physical ranges.

Description

This module defines two functions, add_position_actuator and add_velocity_actuator, that create general actuators with biastype='affine'. The key idea is to map a normalized control range (typically [-1, 1]) to a desired physical range (joint position limits or velocity bounds) via an affine transformation computed from the slope and intercept of the mapping.

For add_position_actuator, the gain and bias parameters are computed as: g0 = kp * slope, b0 = kp * (qposrange[0] - slope * ctrlrange[0]), b1 = -kp, b2 = 0, where slope = (qposrange[1] - qposrange[0]) / (ctrlrange[1] - ctrlrange[0]). This is equivalent to MuJoCo's built-in <position> actuator with a pre-applied affine control transformation.

For add_velocity_actuator, the computation is analogous but with velocity-specific bias: b1 = 0, b2 = -kv. Both functions validate that the target element is a joint, tendon, or site, and reject disallowed keyword arguments that would conflict with the computed parameters (such as biastype, gainprm, biasprm, ctrllimited, and explicit target specifications).

By normalizing control ranges to [-1, 1], these actuators simplify policy learning: the agent always operates in the same control space regardless of the physical joint limits, which is a standard technique in reinforcement learning for locomotion.

Usage

Use these functions when building position-controlled or velocity-controlled walkers that need normalized control inputs. They are particularly important for the CMU humanoid walkers and any walker that uses position or velocity servos with joint-limit-aware scaling.

Code Reference

Source Location

• Repository: Google_deepmind_Dm_control
• File: dm_control/locomotion/walkers/scaled_actuators.py
• Lines: 1-128

Signature

def add_position_actuator(target, qposrange, ctrlrange=(-1, 1),
                          kp=1.0, **kwargs):

def add_velocity_actuator(target, qvelrange, ctrlrange=(-1, 1),
                          kv=1.0, **kwargs):

Import

from dm_control.locomotion.walkers.scaled_actuators import add_position_actuator
from dm_control.locomotion.walkers.scaled_actuators import add_velocity_actuator

I/O Contract

Inputs (add_position_actuator)

Name	Type	Required	Description
target	mjcf element	Yes	A PyMJCF joint, tendon, or site element to be controlled
qposrange	sequence of two floats	Yes	Allowed range of target position [min, max]
ctrlrange	sequence of two floats	No	Allowed range of the control signal (default: (-1, 1))
kp	float	No	Gain parameter for the position actuator (default: 1.0)
**kwargs	dict	No	Additional MJCF actuator attributes (excluding disallowed ones)

Inputs (add_velocity_actuator)

Name	Type	Required	Description
target	mjcf element	Yes	A PyMJCF joint, tendon, or site element to be controlled
qvelrange	sequence of two floats	Yes	Allowed range of target velocity [min, max]
ctrlrange	sequence of two floats	No	Allowed range of the control signal (default: (-1, 1))
kv	float	No	Gain parameter for the velocity actuator (default: 1.0)
**kwargs	dict	No	Additional MJCF actuator attributes (excluding disallowed ones)

Outputs

Name	Type	Description
return	mjcf actuator element	The created general actuator element added to the MJCF model

Usage Examples

from dm_control.locomotion.walkers.scaled_actuators import (
    add_position_actuator, add_velocity_actuator)

# Add a position actuator to a joint with normalized control
joint = walker.mjcf_model.find('joint', 'knee')
pos_actuator = add_position_actuator(
    target=joint,
    qposrange=(-1.57, 0.0),  # joint limits in radians
    ctrlrange=(-1, 1),        # normalized control range
    kp=50.0,                  # proportional gain
    name='knee_pos',
)

# Add a velocity actuator to a joint
vel_actuator = add_velocity_actuator(
    target=joint,
    qvelrange=(-5.0, 5.0),   # velocity limits in rad/s
    ctrlrange=(-1, 1),
    kv=10.0,                  # velocity gain
    name='knee_vel',
)