GreenVLA-5b-stride-1-R2-bridge

RL-Aligned VLA for Bridge (WidowX)

Sber Robotics Center · Manipulation Team

Overview

GreenVLA-5b-stride-1-R2-bridge is the R2 (RL-aligned) checkpoint of the Green-VLA family, fine-tuned on the Bridge dataset for the WidowX robot arm joined with additional trajectories collected in SimplerEnv Bridge environments. Trajectory collection and RL fine-tuning were conducted according to the Trajectory Optimization approach described in the technical report.

Starting from GreenVLA-5b-base-stride-1, this model went through both R1 (supervised fine-tuning) and R2 (RL policy alignment) stages, resulting in significant performance gains over behavior cloning alone.

Evaluation

Evaluated on SimplerEnv WidowX (Bridge) benchmark.

Note: Bridge benchmark results can vary up to ±6% between runs. We recommend averaging over multiple evaluation runs for reliable comparisons.

Partial Success Rate

Task	Success Rate
Put Spoon on Towel	90.6%
Put Carrot on Plate	89.6%
Stack Blocks	99.0%
Put Eggplant in Basket	99.0%
Average	94.5%

Entire Success Rate

Task	Success Rate
Put Spoon on Towel	80.2%
Put Carrot on Plate	76.1%
Stack Blocks	70.8%
Put Eggplant in Basket	94.8%
Average	80.5%

Training

	Details
Base checkpoint	GreenVLA-5b-base-stride-1
Stage	R2 — RL policy alignment
Method	Trajectory optimization (SFT + RL on collected trajectories)
Dataset	IPEC-COMMUNITY/bridge_orig_lerobot + SimplerEnv rollouts
Robot	WidowX (Bridge)
Parameters	~5B

Quick Start

Installation

git clone https://github.com/greenvla/GreenVLA.git
cd GreenVLA
uv sync  # or: pip install -e .

Inference

import numpy as np
import torch
from lerobot.common.policies.factory import load_pretrained_policy
from lerobot.common.utils.torch_observation import (
    move_dict_to_batch_for_inference,
    torch_preprocess_dict_inference,
)

# 1. Load policy and transforms.
policy, input_transforms, output_transforms = load_pretrained_policy(
    "SberRoboticsCenter/GreenVLA-5b-stride-1-R2-bridge",
    data_config_name="bridge",
)
policy.to("cuda").eval()

# 2. Build an observation (replace with real sensor data).
raw_obs = {
    "observation/state": np.random.rand(8).astype(np.float32),  # x y z roll pitch yaw _pad_ gripper
    "observation/image": np.random.randint(0, 256, size=(224, 224, 3), dtype=np.uint8),
    "prompt": "pick up the green block and place it on the plate",
}

# 3. Transform, preprocess, and batch.
obs = input_transforms(raw_obs)
obs = torch_preprocess_dict_inference(obs)
batch = move_dict_to_batch_for_inference(obs, device="cuda")

# 4. Predict actions and post-process.
with torch.inference_mode():
    raw_actions = policy.select_action(batch).cpu().numpy()

actions = output_transforms(
    {"actions": raw_actions, "state": batch["state"].cpu().numpy()}
)["actions"]
# actions shape: (action_horizon, 7) — [x, y, z, roll, pitch, yaw, gripper]

See examples/example_inference_bridge.py for the full runnable script with argument parsing.

Citation

@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},
}