Implementation:Google deepmind Mujoco CtrlNoise Pattern

Knowledge Sources	MuJoCo
Domains	Benchmarking, Stochastic_Processes, Control
Last Updated	2026-02-15 06:00 GMT

Overview

Concrete pattern for generating Ornstein-Uhlenbeck control noise from the MuJoCo testspeed benchmark sample.

Description

The CtrlNoise function in sample/testspeed.cc pre-generates a full control sequence using the OU process driven by Halton quasi-random samples. It produces a flat array of nsteps * nu control values that are applied sequentially during the benchmark rollout.

Usage

Use as a pattern for generating benchmark control inputs. Customize noise amplitude (ctrl_noise_std) and rate (ctrl_noise_rate) for different model types.

Code Reference

Source Location

• Repository: mujoco
• File: sample/testspeed.cc
• Lines: 64-101

Signature

// Local helper function (not part of public API)
std::vector<mjtNum> CtrlNoise(
    const mjModel* m,
    int nsteps,
    mjtNum ctrl_noise_std,   // default 1.0
    mjtNum ctrl_noise_rate,  // default 0.01
    int key                  // keyframe index for initialization
);

Import

// Not a library function - pattern from sample/testspeed.cc
#include <mujoco/mujoco.h>
#include <vector>

I/O Contract

Inputs

Name	Type	Required	Description
m	const mjModel*	Yes	Model (for nu actuator count and keyframes)
nsteps	int	Yes	Number of timesteps to generate
ctrl_noise_std	mjtNum	No	Noise amplitude (default 1.0)
ctrl_noise_rate	mjtNum	No	Mean-reversion rate (default 0.01)
key	int	No	Keyframe index for control initialization

Outputs

Name	Type	Description
return	std::vector<mjtNum>	Pre-generated control sequence of length nsteps * m->nu

Usage Examples

#include <mujoco/mujoco.h>
#include <vector>

mjModel* m = mj_loadXML("humanoid.xml", NULL, error, 1000);
mjData* d = mj_makeData(m);

// Generate 10000 steps of control noise
std::vector<mjtNum> ctrl = CtrlNoise(m, 10000, 1.0, 0.01, -1);

// Apply during simulation
for (int i = 0; i < 10000; i++) {
    mju_copy(d->ctrl, ctrl.data() + i * m->nu, m->nu);
    mj_step(m, d);
}

Related Pages

Implements Principle

• Principle:Google_deepmind_Mujoco_Control_Noise_Generation