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:

Stage	Name	Purpose
Base	Multi-Embodiment Pretrained Model	VLM backbone with multimodal grounding, trained on 3,000+ hours of demonstrations
R1	Embodiment-Specific Adaptation	Supervised fine-tuning for target robot
R2	RL Policy Alignment	Trajectory optimization beyond behavior cloning

For full details, see our paper and project page.

---

Models

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

Model	Stage	Params	Description	Link
GreenVLA-2b-base	Base	2B	Base pretrained model (lightweight)	Hub
GreenVLA-5b-base-stride-1	Base	5B	Base pretrained, action expert depth = VLM depth	Hub
GreenVLA-5b-base-stride-4	Base	5B	Base pretrained, action expert depth = VLM depth / 4	Hub
GreenVLA-5b-stride-1-R1-bridge	R1	5B	Fine-tuned on Bridge (WidowX)	Hub
GreenVLA-5b-stride-1-R2-bridge	R2	5B	RL-aligned on Bridge (WidowX)	Hub
GreenVLA-5b-stride-4-R1-fractal	R1	5B	Fine-tuned on Fractal (Google Robot)	Hub
GreenVLA-5b-stride-4-R2-calvin	R2	5B	RL-aligned on CALVIN	Hub

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:

Criterion	Stride-1	Stride-4
Inference VRAM	~12.5 GB	~11.1 GB
Training batch size (80 GB GPU, 3×448² images, tokenizer_max_length=640)	5	6
Inference speed (4B proxy, see below)	Slower	Faster

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

Compiled	Stride-1	Stride-4
No	0.273 s	0.124 s
Yes	0.181 s	0.098 s

• 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

Using uv (recommended)

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

Using micromamba / conda

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)

Model	Visual Matching	Variant Agg.	Overall
Green-VLA stride-4 R1	77.0%	66.7%	71.8%

SimplerEnv — WidowX (Bridge)

Model	Partial Avg	Entire Avg
Green-VLA stride-1 R2	94.5%	80.5%
Green-VLA stride-1 R1	89.6%	72.9%

CALVIN

Model	Avg Chain Length
Green-VLA stride-4 R2	4.57

---

Documentation

Guide	Description
Fine-Tuning	Dataset statistics, configuration, training
Inference	Loading models, running inference, example notebooks

---

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.

© 2026 Sber Robotics Center · Manipulation Team