Commit ·
c2dfb89
0
Parent(s):
Initial release
Browse files
README.md
ADDED
|
@@ -0,0 +1,182 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: lerobot
|
| 3 |
+
license: apache-2.0
|
| 4 |
+
language:
|
| 5 |
+
- en
|
| 6 |
+
base_model:
|
| 7 |
+
- SberRoboticsCenter/GreenVLA-5b-base
|
| 8 |
+
pipeline_tag: robotics
|
| 9 |
+
tags:
|
| 10 |
+
- robotics
|
| 11 |
+
- vla
|
| 12 |
+
- vision-language-action
|
| 13 |
+
- manipulation
|
| 14 |
+
- flow-matching
|
| 15 |
+
- action-prediction
|
| 16 |
+
- green-vla
|
| 17 |
+
- bridge
|
| 18 |
+
- widowx
|
| 19 |
+
datasets:
|
| 20 |
+
- IPEC-COMMUNITY/bridge_orig_lerobot
|
| 21 |
+
model-index:
|
| 22 |
+
- name: GreenVLA-5b-R1-bridge
|
| 23 |
+
results:
|
| 24 |
+
- task:
|
| 25 |
+
type: robotics
|
| 26 |
+
name: SimplerEnv WidowX (Bridge)
|
| 27 |
+
dataset:
|
| 28 |
+
type: IPEC-COMMUNITY/bridge_orig_lerobot
|
| 29 |
+
name: Bridge
|
| 30 |
+
metrics:
|
| 31 |
+
- type: success_rate
|
| 32 |
+
name: Partial Average
|
| 33 |
+
value: 86.5
|
| 34 |
+
- type: success_rate
|
| 35 |
+
name: Entire Average
|
| 36 |
+
value: 71.9
|
| 37 |
+
---
|
| 38 |
+
|
| 39 |
+
<div align="center">
|
| 40 |
+
|
| 41 |
+
# GreenVLA-5b-R1-bridge
|
| 42 |
+
|
| 43 |
+
### Embodiment-Adapted VLA for Bridge (WidowX)
|
| 44 |
+
|
| 45 |
+
**Sber Robotics Center · Manipulation Team**
|
| 46 |
+
|
| 47 |
+
[](https://arxiv.org/abs/2602.00919)
|
| 48 |
+
[](https://greenvla.github.io/)
|
| 49 |
+
[](https://github.com/greenvla/GreenVLA)
|
| 50 |
+
|
| 51 |
+
</div>
|
| 52 |
+
|
| 53 |
+
---
|
| 54 |
+
|
| 55 |
+
## Overview
|
| 56 |
+
|
| 57 |
+
**GreenVLA-5b-R1-bridge** is the R1 (embodiment-adapted) checkpoint of the [Green-VLA](https://arxiv.org/abs/2602.00919) family, fine-tuned on the [Bridge](https://huggingface.co/datasets/IPEC-COMMUNITY/bridge_orig_lerobot) dataset for the WidowX robot arm.
|
| 58 |
+
|
| 59 |
+
Starting from the [GreenVLA-5b-base](https://huggingface.co/SberRoboticsCenter/GreenVLA-5b-base) pretrained checkpoint, this model was adapted via supervised fine-tuning (R1 stage) to the Bridge embodiment, achieving strong manipulation performance on the SimplerEnv benchmark.
|
| 60 |
+
|
| 61 |
+
## Evaluation
|
| 62 |
+
|
| 63 |
+
Evaluated on **SimplerEnv WidowX (Bridge)** benchmark with default episode length:
|
| 64 |
+
|
| 65 |
+
### Partial Success Rate
|
| 66 |
+
|
| 67 |
+
| Task | Success Rate |
|
| 68 |
+
|------|:---:|
|
| 69 |
+
| Put Spoon on Towel | 87.5% |
|
| 70 |
+
| Put Carrot on Plate | 83.3% |
|
| 71 |
+
| Stack Blocks | 79.2% |
|
| 72 |
+
| Put Eggplant in Basket | 95.8% |
|
| 73 |
+
| **Average** | **86.5%** |
|
| 74 |
+
|
| 75 |
+
### Entire Success Rate
|
| 76 |
+
|
| 77 |
+
| Task | Success Rate |
|
| 78 |
+
|------|:---:|
|
| 79 |
+
| Put Spoon on Towel | 79.2% |
|
| 80 |
+
| Put Carrot on Plate | 70.8% |
|
| 81 |
+
| Stack Blocks | 41.7% |
|
| 82 |
+
| Put Eggplant in Basket | 95.8% |
|
| 83 |
+
| **Average** | **71.9%** |
|
| 84 |
+
|
| 85 |
+
## Training
|
| 86 |
+
|
| 87 |
+
| | Details |
|
| 88 |
+
|---|---|
|
| 89 |
+
| **Base checkpoint** | [GreenVLA-5b-base](https://huggingface.co/SberRoboticsCenter/GreenVLA-5b-base) |
|
| 90 |
+
| **Stage** | R1 — Embodiment-specific adaptation |
|
| 91 |
+
| **Method** | Supervised fine-tuning |
|
| 92 |
+
| **Dataset** | [IPEC-COMMUNITY/bridge_orig_lerobot](https://huggingface.co/datasets/IPEC-COMMUNITY/bridge_orig_lerobot) |
|
| 93 |
+
| **Robot** | WidowX (Bridge) |
|
| 94 |
+
| **Parameters** | ~5B |
|
| 95 |
+
|
| 96 |
+
## Quick Start
|
| 97 |
+
|
| 98 |
+
### Installation
|
| 99 |
+
|
| 100 |
+
```bash
|
| 101 |
+
git clone https://github.com/greenvla/GreenVLA.git
|
| 102 |
+
cd GreenVLA
|
| 103 |
+
uv sync # or: pip install -e .
|
| 104 |
+
```
|
| 105 |
+
|
| 106 |
+
### Inference
|
| 107 |
+
|
| 108 |
+
```python
|
| 109 |
+
import numpy as np
|
| 110 |
+
import torch
|
| 111 |
+
from lerobot.common.policies.factory import load_pretrained_policy
|
| 112 |
+
from lerobot.common.utils.torch_observation import (
|
| 113 |
+
move_dict_to_batch_for_inference,
|
| 114 |
+
torch_preprocess_dict_inference,
|
| 115 |
+
)
|
| 116 |
+
|
| 117 |
+
# 1. Load policy and transforms.
|
| 118 |
+
policy, input_transforms, output_transforms = load_pretrained_policy(
|
| 119 |
+
"SberRoboticsCenter/GreenVLA-5b-R1-bridge",
|
| 120 |
+
data_config_name="bridge",
|
| 121 |
+
)
|
| 122 |
+
policy.to("cuda").eval()
|
| 123 |
+
|
| 124 |
+
# 2. Build an observation (replace with real sensor data).
|
| 125 |
+
raw_obs = {
|
| 126 |
+
"observation/state": np.random.rand(8).astype(np.float32), # x y z roll pitch yaw _pad_ gripper
|
| 127 |
+
"observation/image": np.random.randint(0, 256, size=(224, 224, 3), dtype=np.uint8),
|
| 128 |
+
"prompt": "pick up the green block and place it on the plate",
|
| 129 |
+
}
|
| 130 |
+
|
| 131 |
+
# 3. Transform, preprocess, and batch.
|
| 132 |
+
obs = input_transforms(raw_obs)
|
| 133 |
+
obs = torch_preprocess_dict_inference(obs)
|
| 134 |
+
batch = move_dict_to_batch_for_inference(obs, device="cuda")
|
| 135 |
+
|
| 136 |
+
# 4. Predict actions and post-process.
|
| 137 |
+
with torch.inference_mode():
|
| 138 |
+
raw_actions = policy.select_action(batch).cpu().numpy()
|
| 139 |
+
|
| 140 |
+
actions = output_transforms(
|
| 141 |
+
{"actions": raw_actions, "state": batch["state"].cpu().numpy()}
|
| 142 |
+
)["actions"]
|
| 143 |
+
# actions shape: (action_horizon, 7) — [x, y, z, roll, pitch, yaw, gripper]
|
| 144 |
+
```
|
| 145 |
+
|
| 146 |
+
See [`examples/example_inference_bridge.py`](https://github.com/greenvla/GreenVLA/blob/main/examples/example_inference_bridge.py) for the full runnable script with argument parsing.
|
| 147 |
+
|
| 148 |
+
## Model Family
|
| 149 |
+
|
| 150 |
+
| Model | Stage | Params | Description | Link |
|
| 151 |
+
|-------|:-----:|:------:|-------------|:----:|
|
| 152 |
+
| **GreenVLA-2b-base** | Base | 2B | Base pretrained (lightweight) | [Hub](https://huggingface.co/SberRoboticsCenter/GreenVLA-2b-base) |
|
| 153 |
+
| **GreenVLA-5b-base** | Base | 5B | Base pretrained (recommended) | [Hub](https://huggingface.co/SberRoboticsCenter/GreenVLA-5b-base) |
|
| 154 |
+
| **GreenVLA-5b-R1-bridge** | R1 | 5B | Fine-tuned on Bridge (WidowX) | You are here |
|
| 155 |
+
| **GreenVLA-5b-R2-bridge** | R2 | 5B | RL-aligned on Bridge (WidowX) | [Hub](https://huggingface.co/SberRoboticsCenter/GreenVLA-5b-R2-bridge) |
|
| 156 |
+
| **GreenVLA-5b-R1-fractal** | R1 | 5B | Fine-tuned on Fractal (Google Robot) | [Hub](https://huggingface.co/SberRoboticsCenter/GreenVLA-5b-R1-fractal) |
|
| 157 |
+
|
| 158 |
+
## Citation
|
| 159 |
+
|
| 160 |
+
```bibtex
|
| 161 |
+
@misc{greenvla,
|
| 162 |
+
title = {Green-VLA: Staged Vision-Language-Action Model for Generalist Robots},
|
| 163 |
+
author = {I. Apanasevich and M. Artemyev and R. Babakyan and P. Fedotova and
|
| 164 |
+
D. Grankin and E. Kupryashin and A. Misailidi and D. Nerus and
|
| 165 |
+
A. Nutalapati and G. Sidorov and I. Efremov and M. Gerasyov and
|
| 166 |
+
D. Pikurov and Y. Senchenko and S. Davidenko and D. Kulikov and
|
| 167 |
+
M. Sultankin and K. Askarbek and O. Shamanin and D. Statovoy and
|
| 168 |
+
E. Zalyaev and I. Zorin and A. Letkin and E. Rusakov and
|
| 169 |
+
A. Silchenko and V. Vorobyov and S. Sobolnikov and A. Postnikov},
|
| 170 |
+
year = {2026},
|
| 171 |
+
eprint = {2602.00919},
|
| 172 |
+
archivePrefix = {arXiv},
|
| 173 |
+
primaryClass = {cs.RO},
|
| 174 |
+
url = {https://arxiv.org/abs/2602.00919},
|
| 175 |
+
}
|
| 176 |
+
```
|
| 177 |
+
|
| 178 |
+
<div align="center">
|
| 179 |
+
|
| 180 |
+
© 2026 Sber Robotics Center · Manipulation Team
|
| 181 |
+
|
| 182 |
+
</div>
|