Skip to content
RobothecaBeta

AI model

Perceptive Humanoid Parkour (PHP)

By Amazon FAR, United States (checked on github.com/amazon-far (opens in a new tab)), from github.com/amazon-far/php_parkour (opens in a new tab)

  • Humanoids and robot learning
  • Apache-2.0 (code)
  • Apache-2.0 (weights)

A framework that lets humanoid robots perform long-horizon, vision-based parkour across obstacle courses by chaining retargeted human skills via motion matching.

Open on GitHub (opens in a new tab)

Facts

Code last updated
13 Oct 2026Newest commit on main
Maintenance
Activea commit in the last 90 days
Difficulty
Advanced
Hardware it needs, as the project states it
Motion matching needs Conda and no IsaacSim; training and evaluation need Linux with an NVIDIA GPU and IsaacSim; sim2sim needs Holosoma's MuJoCo and inference environments.

Checked on GitHub on 6 Nov 2026.

Licenses

Code
Apache-2.0 (opens in a new tab)

Read by Robotheca from LICENSE (opens in a new tab) (Apache-2.0) at commit c7c9ce3 (opens in a new tab), on 4 Nov 2026.

Weights
Apache-2.0 (opens in a new tab)

Read on github.com/amazon-far/php_parkour/releases/tag/student-asse… (opens in a new tab).

Commercial use
Yes

Overview

What it does: Robotheca's summary

  • Composes retargeted atomic human skills into long-horizon trajectories offline using motion matching.
  • Trains RL expert policies to track those trajectories using privileged state and terrain observations.
  • Distills the experts into a depth-based, multi-skill student policy using DAgger and RL.
  • Provides an example student policy that can be run in the MuJoCo simulator via exported ONNX models.
  • Includes guides for motion matching, training and evaluation, and sim2sim deployment.

Written by Robotheca from the project's README (opens in a new tab) at commit c7c9ce3, not by its creator.

The README

The creator's own README, copied as written: README.md by Amazon FAR, at commit c7c9ce3, under Apache-2.0 (opens in a new tab). The original on GitHub (opens in a new tab).

Read the README here

[RSS 2026] Perceptive Humanoid Parkour: Chaining Dynamic Human Skills via Motion Matching

Zhen Wu*, Xiaoyu Huang*, Lujie Yang*, Yuanhang Zhang, Xi Chen, Pieter Abbeel†, Rocky Duan†, Angjoo Kanazawa†, Carmelo Sferrazza†, Guanya Shi†, C. Karen Liu†

Project page (opens in a new tab) · arXiv (opens in a new tab) · Video (opens in a new tab)

Parkour Demo

Overview

PHP (Perceptive Humanoid Parkour) enables humanoid robots to autonomously perform long-horizon, vision-based parkour across challenging obstacle courses. The framework has three steps:

  1. Motion matching. Compose retargeted atomic human skills into long-horizon trajectories offline using motion matching.
  2. Teacher training. Train RL experts to track the trajectories with privileged state and terrain observations.
  3. Student training. Distill the experts into a depth-based, multi-skill policy using DAgger + RL.
Get started

For sim2sim with an existing ONNX pair, use the PHP launchers and pinned Holosoma setup (opens in a new tab).

For motion matching and training, clone PHP with its pinned Holosoma submodule:

git clone --recurse-submodules https://github.com/amazon-far/php_parkour.git
cd php_parkour

Choose a guide below for environment setup and commands. For training, use the prepared example datasets (opens in a new tab) or generate your own with motion matching. If you already have an exported ONNX pair, go directly to sim2sim.

GuideWorkflowRequirements
Motion matching (opens in a new tab)Generate and inspect motion/terrain datasets; upload them for training or add new skillsConda; no IsaacSim
Training and evaluation (opens in a new tab)Train teachers, distill a student, evaluate checkpoints, and export ONNXLinux/NVIDIA with IsaacSim
Sim2sim (opens in a new tab)Run the exported depth policy in MuJoCoHolosoma's MuJoCo and inference environments
🏃 Try the demo

PHP depth locomotion demo in MuJoCo

To try the example student policy in MuJoCo, download the validated ONNX pair from the student release (opens in a new tab). Run from the PHP checkout root:

python scripts/download_assets.py student

The release contains sanitized depth_backbone.onnx and student.onnx files, without raw .pt or teacher checkpoints. Files default to ~/.cache/php-parkour/student; pass --destination DIR to choose another directory. The downloader verifies the archive and both model files against release-assets.json (opens in a new tab).

Follow the sim2sim guide (opens in a new tab) to run the policy. You can also use the automatically exported ONNX pair (opens in a new tab) from your own student training run.

Community projects
Citation
@article{wu2026perceptive,
  title={Perceptive humanoid parkour: Chaining dynamic human skills via motion matching},
  author={Wu, Zhen and Huang, Xiaoyu and Yang, Lujie and Zhang, Yuanhang and Chen, Xi and Abbeel, Pieter and Duan, Rocky and Kanazawa, Angjoo and Sferrazza, Carmelo and Shi, Guanya and others},
  journal={arXiv preprint arXiv:2602.15827},
  year={2026}
}

Related: OmniRetarget (opens in a new tab).

The model card

Read the README on GitHub (opens in a new tab)

Files and versions

Facts read at commit c7c9ce3

Robotheca never hosts the code or the weights: each link goes to the creator's own page.

Use it

How to load it, in the creator's words: The project's website (opens in a new tab)

Robotheca doesn't review code. Read it before running it.

Robotheca lists AI models as a reference: nothing is hosted or run here, and every fact links to where it was read.