X-Square-Robot/wall-x

What is it?

Key features

• 01Fine-tunes models on LeRobot-format datasets using an FSDP trainer.
• 02Supports flow-matching and FAST action generation branches.
• 03Loads pretrained WALL-OSS checkpoints hosted on Hugging Face.
• 04Serves model actions over a WebSocket server.
• 05Includes open-loop evaluation against LeRobot ground-truth data and LIBERO simulator launchers.

Use cases

• Fine-tuning a vision-language-action model on a custom robot manipulation dataset.
• Running inference with a pretrained embodied model on a real robot or in simulation.
• Comparing open-loop action predictions against recorded demonstrations.
• Researching flow-matching or FAST action heads for robot policies.

Quick start

1. 1

   Create and activate a conda environment with Python 3.10

   conda create --name wallx python=3.10
2. 2

   Install the requirements

   pip install -r requirements.txt
3. 3

   Install Wall-X, which compiles the CUDA operators

   MAX_JOBS=8 pip install --no-build-isolation -e .
4. 4

   Run the minimal inference example with a downloaded checkpoint

   python scripts/fake_inference.py --checkpoint-path <path/to/checkpoint>

Things to know

• !Requires an NVIDIA CUDA GPU, and the install compiles CUDA operators and flash-attn.
• !Setup is heavy, with several dependencies such as LeRobot, DMuon and flash-attn to install by hand.
• !The README targets Ubuntu 22.04 with CUDA 12.x and Python 3.10.

In collections

Topics

FAQ

What is wall-x?

Training and inference code for the WALL-OSS series of open-source embodied foundation models for general-purpose robot manipulation. Wall-X is the training and inference codebase for the WALL series of embodied foundation models, which aim to give robots general-purpose manipulation skills learned from physical interaction data. It covers the full workflow from preparing LeRobot-format datasets and configuring models to fine-tuning, running inference, and serving a model for closed-loop or open-loop evaluation.

Is wall-x free to use? What license does it have?

wall-x is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.

Is wall-x still maintained?

Updated within 6 months. The last commit to wall-x was on Sep 18, 2026.

What can I use wall-x for?

Fine-tuning a vision-language-action model on a custom robot manipulation dataset. Running inference with a pretrained embodied model on a real robot or in simulation. Comparing open-loop action predictions against recorded demonstrations. Researching flow-matching or FAST action heads for robot policies.

How do I install wall-x?

Create and activate a conda environment with Python 3.10: conda create --name wallx python=3.10 Install the requirements: pip install -r requirements.txt Install Wall-X, which compiles the CUDA operators: MAX_JOBS=8 pip install --no-build-isolation -e . Run the minimal inference example with a downloaded checkpoint: python scripts/fake_inference.py --checkpoint-path <path/to/checkpoint>

Where can I get wall-x?

The source code and installation instructions are on GitHub: https://github.com/X-Square-Robot/wall-x