Open Source RoboticsRobots, software & learning

Universal Humanoid Control

Software

MuJoCo human-motion imitation controller using SMPL-family humanoids and residual force control, with pretrained evaluation and training workflows. [1]

Installation & first result →Official websiteGetting started

For
Researchers reproducing legacy human-motion imitation experiments [1]
Source rights
Some licenses are unresolved. Review licenses
Unknowns
operating system, compute requirements not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
License
BSD-3-Clause-Clear · license not fully assessed for model weights, datasets · Details by part

Requirements · Research details and dates

On this page

Overview

Universal Humanoid Control (UHC) trains a simulated humanoid to follow reference human motion. It supports SMPL-family body models and relies heavily on residual force control to help stabilize imitation. Those extra simulated forces matter when interpreting a result: successful tracking is not evidence that a physical humanoid can execute the same motion. [1]

Is this for you?

Known limitations
  • Residual force assistance is part of the controller and is not a physical-robot actuator capability.
  • The evaluation entry point selects CPU despite GPU-oriented setup instructions.
  • Dependencies include an unpinned smplx fork at master.

Editorial assessment from cited documentation; no runtime testing.

[1] [2] [3] · Editorial assessment

What you need

Dependencies
  • Python 3.8 setup
  • MuJoCo 2.1.0 Linux x86_64 distribution
  • mujoco-py >=2.1,<2.2
  • PyTorch setup with CUDA 11.6
  • External body models and checkpoints

[1] [2]

Installation & first result

Open the upstream getting-started guide [1]

  • Prepare the documented Python 3.8 and MuJoCo 2.1.0 environment.
  • Place separately licensed SMPL-family assets in data/smpl and obtain the pretrained checkpoint.
  • Run the implicit-controller evaluation command in the overview.

Editorial assessment from cited documentation; no runtime testing.

[1] · Editorial assessment

Installation

Follow the legacy MuJoCo/Python setup, acquire body models and the matching pretrained controller, then run the configured evaluation.

[1]

Hardware, models & data

Inputs and outputs

Reference human-motion frames drive a simulated humanoid; evaluation provides visualization or saved statistics.

[1] [3]

Project notes

Reproduce the older simulator environment deliberately

This profile reviews commit 2419deef54958b47b0a0cbf0aee50c82426e7950, with no GitHub release listed. Upstream uses Python 3.8, a PyTorch setup with CUDA 11.6 and the Linux x86_64 MuJoCo 2.1.0 distribution. Requirements specify mujoco-py>=2.1,<2.2, not the current native MuJoCo Python interface. They also install a separate smplx fork from an unpinned master branch. Record resolved versions before comparing results. [1] [2] [6]

Obtain the relevant SMPL-family body files under their own terms, place them in data/smpl with upstream's expected names, and acquire the pretrained controller using its link or bash download_data.sh. The implicit controller example is python scripts/eval_uhc.py --cfg uhc_implicit --epoch 19000 --data sample_data/amass_copycat_take5_test_small.pkl. Use the matching checkpoint/configuration pair; the shape-varying and explicit-force examples use different epochs. [1] [4]

The evaluation script defaults to visualization. Add --mode stats for the statistics route. Despite the GPU-oriented environment instructions, this entry point explicitly sets its policy device to CPU; it also sets the default tensor dtype to float64. That is source behavior, not a claim about runtime speed or the requirements of other training scripts. [3]

Distinguish this implementation from its successors

The README points to PHC for motion tracking without residual force control, and to PHC-MJX and SMPLSim for newer MuJoCo work. These are separate projects, not drop-in package upgrades established by this review. UHC remains useful for reproducing its own experiments and studying the residual-force formulation. [1]

Repository source carries BSD-3-Clause-Clear terms. Pretrained checkpoints, SMPL-family assets and AMASS data are separate artifacts whose rights are not settled by that source license. Their download routes are documented, but external payloads were not acquired here. No installation, evaluation, training or physical control was tested. [1] [7]

What we checked

Source researched only; no installation, training, benchmark or physical hardware test performed. No affiliate links or reused media.

Not yet established

  • Requirements: operating system, host computer
  • Version-specific hardware compatibility (ecosystem tags alone do not establish it)

Updates

Upstream last updated
Feb 24, 2024 [6]
Latest release
Not established
Profile updated
Oct 17, 2026
Sources last checked
Oct 17, 2026
Next review
Nov 17, 2026
Reviewed revision
commit snapshot; no GitHub releases listed / 2419deef54958b47b0a0cbf0aee50c82426e7950 · software repository

Source files and licenses

Source-researched legacy simulation controller with residual-force and external-artifact scope disclosed.

LayerAvailability / license
softwareavailable · BSD-3-Clause-Clear
Repository materials only; external components retain their own terms. [7]
model weightsavailable · License not fully assessed
Download routes documented; external checkpoint/body-model/data payloads not acquired and artifact-specific terms unresolved. [1]
datasetsavailable · License not fully assessed
Download routes documented; external checkpoint/body-model/data payloads not acquired and artifact-specific terms unresolved. [1]
documentationavailable · BSD-3-Clause-Clear
Repository materials only; external components retain their own terms. [7]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

  • S001 Controller setup and scope
    Revision, date & where cited

    ZhengyiLuo/UniversalHumanoidControl contributors · github.com · revision 2419dee · Accessed Oct 17, 2026


    Used for: Overview, Is this for you?, What you need, Installation & first result and 4 more
  • S002 Dependencies
    Revision, date & where cited

    ZhengyiLuo/UniversalHumanoidControl contributors · github.com · revision 2419dee · Accessed Oct 17, 2026


    Used for: Is this for you?, What you need, Reproduce the older simulator environment deliberately
  • S003 Evaluation entry point
    Revision, date & where cited

    ZhengyiLuo/UniversalHumanoidControl contributors · github.com · revision 2419dee · Accessed Oct 17, 2026


    Used for: Is this for you?, Hardware, models & data, Reproduce the older simulator environment deliberately
  • S004 Implicit controller configuration
    Revision, date & where cited

    ZhengyiLuo/UniversalHumanoidControl contributors · github.com · revision 2419dee · Accessed Oct 17, 2026


    Used for: Reproduce the older simulator environment deliberately
  • S007 Source license
    Revision, date & where cited

    ZhengyiLuo/UniversalHumanoidControl contributors · github.com · revision 2419dee · Accessed Oct 17, 2026


    Used for: Distinguish this implementation from its successors, Source files and licenses
  • S006 Reviewed release and revision
    Revision, date & where cited

    ZhengyiLuo/UniversalHumanoidControl / GitHub · github.com · revision 2419dee · commit snapshot; no GitHub releases listed / · Accessed Oct 17, 2026


    Used for: Reproduce the older simulator environment deliberately, What we checked

Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5

Change history