Green-VLA

Staged Vision-Language-Action Model for Generalist Robots

Sber Robotics Center · Manipulation Team

Paper | Project Page | Models | Quick Start | Benchmarks

Overview

Green-VLA is a ~4B-parameter Vision-Language-Action model built on a staged curriculum:

For full details, see our paper and project page.

Models

We release model checkpoints spanning base pretrained models and embodiment-adapted variants:

Recommendation: Start with GreenVLA-5b-base-stride-1 for fine-tuning on your own embodiment, or use one of the R1/R2 checkpoints for direct evaluation.

Choosing stride-1 vs stride-4 (5B base)

The 5B base is available in two action-expert variants: stride-1 (action expert has the same number of layers as the VLM) and stride-4 (action expert has 4× fewer layers). Use the following to decide:

Inference time (mean seconds per forward, 4B model; 5 warmup + 50 benchmark iterations):

• Prefer stride-1 when you need maximum action capacity and have enough VRAM; use it for Bridge/CALVIN and when fine-tuning from scratch.
• Prefer stride-4 when you are memory- or latency-bound; we release a Fractal R1 checkpoint in this variant.

Quick Start

Installation

git clone https://github.com/greenvla/GreenVLA.git
cd GreenVLA
uv sync

git clone https://github.com/greenvla/GreenVLA.git
cd GreenVLA

micromamba env create -n greenvla python=3.11
micromamba activate greenvla
pip install -e .

Inference

Load a model and build the full inference pipeline in a single call:

from lerobot.common.policies.factory import load_pretrained_policy

policy, input_transforms, output_transforms = load_pretrained_policy(
    "SberRoboticsCenter/GreenVLA-5b-base-stride-1",
    data_config_name="bridge",
)

This downloads the config, weights, and normalization statistics from the Hub automatically. It also works with local checkpoint paths:

policy, input_transforms, output_transforms = load_pretrained_policy(
    "/path/to/checkpoint",
    data_config_name="bridge",
    config_overrides={"device": "cuda:0"},
)

See docs/INFERENCE.md for the full inference guide and example notebooks.

Benchmarks

SimplerEnv — Google Robot (Fractal)

SimplerEnv — WidowX (Bridge)

CALVIN

Documentation

Project Structure

GreenVLA/
├── assets/                  # Images, videos, and other media
├── docs/                    # Detailed guides
│   ├── FINE_TUNING.md       # Fine-tuning guide
│   └── INFERENCE.md         # Inference guide & examples
├── lerobot/
│   ├── conf/                # Hydra configs (policy, dataset, training)
│   ├── common/
│   │   └── policies/        # Policy implementations
│   │       └── greenvla_policy/
│   └── scripts/             # Training & inference scripts
└── examples/                # Inference examples & notebooks

Citation

If you find Green-VLA useful, please cite our paper:

@misc{apanasevich2026greenvlastagedvisionlanguageactionmodel,
    title   = {Green-VLA: Staged Vision-Language-Action Model for Generalist Robots},
    author  = {I. Apanasevich and M. Artemyev and R. Babakyan and P. Fedotova and
               D. Grankin and E. Kupryashin and A. Misailidi and D. Nerus and
               A. Nutalapati and G. Sidorov and I. Efremov and M. Gerasyov and
               D. Pikurov and Y. Senchenko and S. Davidenko and D. Kulikov and
               M. Sultankin and K. Askarbek and O. Shamanin and D. Statovoy and
               E. Zalyaev and I. Zorin and A. Letkin and E. Rusakov and
               A. Silchenko and V. Vorobyov and S. Sobolnikov and A. Postnikov},
    year    = {2026},
    eprint  = {2602.00919},
    archivePrefix = {arXiv},
    primaryClass  = {cs.RO},
    url     = {https://arxiv.org/abs/2602.00919},
}

⚠ Acknowledgements

This project draws inspiration and references from several notable open-source initiatives, including:

• LeRobot
• OpenPI
• StarVLA

The codebase was originally forked from LeRobot.