Vlarlkit: Train vision-language-action models with RL

Understand the project

Practical overview

When it fits

Use Vlarlkit when you need a modifiable PyTorch research codebase for VLA reinforcement learning and can manage the separate model-backend and benchmark environments required by an experiment.

Who it is for

Researchers developing or modifying reinforcement-learning methods for vision-language-action models.
Robotics and embodied-AI teams running VLA experiments with supported model backends and simulation benchmarks.

Documented uses

Build or adapt VLA reinforcement-learning experiments using distinct policy, rollout, runner and model layers.
Run listed on-policy, off-policy or model-based approaches with supported VLA base models.
Collect benchmark data asynchronously alongside model updates for off-policy training.

Before you choose

Requirements & limits

The repository is public and licensed under MIT.

Core installation uses uv; each model backend has its own uv project and dependencies.
Benchmark environments run separately as ZMQ processes with independent dependencies.
The quick-start workflow expects a downloaded SFT model and a configuration entry pointing to its local path.
The repository labels its supported algorithms, models and benchmarks as work in progress.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Set up the core library and prepare a benchmark training run

Install the core package, select the documented LIBERO and OpenPI components, and configure the example on-policy run.

Before you start

Python environment managed with uv
A downloaded SFT checkpoint for the selected benchmark
The Vlarlkit repository

Steps

Clone the repository, enter its directory, run the command below, then run the command below.
Documented command
```
uv sync
```
Documented command
```
uv pip install -e .
```
Prepare the OpenPI backend with the command below and run its documented Transformers patch script.
Documented command
```
uv sync --project model_backends/openpi
```
Install the LIBERO benchmark environment with the command below.
Documented command
```
bash third_party/install_libero.sh
```
Download an appropriate SFT checkpoint, set model.model_path in the example configuration to the local checkpoint path, and launch the command below.
Documented command
```
bash examples/run_onpolicy_rl.sh
```

Expected outcome

A configured example on-policy reinforcement-learning run using the chosen SFT model, OpenPI backend and LIBERO environment.

Guide limitations

The guide follows repository instructions; it is not a verified execution or compatibility test.
The repository marks its supported algorithms, base models and benchmarks as work in progress.
The model backend and benchmark have separate dependencies and setup steps.

Documentation 1 ↗

Evidence and transparency

Sources & review

Official repository ↗

Source record details

Source availability: Active
Sources checked: 13 October 2026
Original website title: VLARLKit/VLARLKit · Official homepage title · retained in its source language

How we review sources Suggest a correction

Vlarlkit