qirui095/Flat-Stove-Turn-On

Dataset Overview

数据集概述

This dataset was created using VLA-Arena [https://github.com/PKU-Alignment/VLA-Arena/tree/main].

This dataset contains Level 0 demonstration data for the "rotate switch to turn on the flat stove" task collected through the VLA-Arena framework, along with Level 0 / Level 1 / Level 2 scene files, used for training and evaluating Vision-Language-Action (VLA) models.

该数据集包含通过 VLA-Arena 框架收集的"旋转开关打开炉台"任务的 Level 0 演示数据，和 Level 0 / Level 1 / Level 2 场景文件，用于训练和评估视觉 - 语言 - 动作（VLA）模型。

Task Description

任务描述

• Task Name: Flat Stove Turn-On (Turn on the flat stove by rotating the switch)
• 任务名称: Flat Stove Turn-On (通过旋转开关打开炉台)
• Task Type: Tabletop Manipulation
• 任务类型: Tabletop Manipulation (桌面操作)
• Difficulty Levels: Level 0, Level 1, Level 2
• 难度级别: Level 0, Level 1, Level 2
• Number of Scenes: 6 scenes per difficulty level (18 scenes total)
• 场景数量: 每个难度级别 6 个场景（共 18 个场景）
• Language Instructions: Multiple expressions describing the same task (e.g., "Turn on the flat stove by rotating the switch", "Activate the flat stove by turning its knob", etc.)
• 语言指令: 使用多种表达方式描述相同的任务（如 "Turn on the flat stove by rotating the switch", "Activate the flat stove by turning its knob" 等）

Scene Characteristics

场景特点

Visual Perturbations
视觉扰动

• Random color variations
• 随机颜色变化
• Image effect adjustments (brightness, contrast, saturation, color temperature)
• 图像效果调整（亮度、对比度、饱和度、色温）
• Enhances model visual robustness
• 增强模型的视觉鲁棒性

Language Perturbations
语言扰动

• Level 2 introduces 6 different linguistic expressions
• Level 2 引入 6 种不同的语言表达
• Tests model understanding of language variation
• 测试模型对语言变化的理解能力

Level 0 - Basic Difficulty

Level 0 - 基础难度

• Stove position varies
• 炉台位置变化
• Some scenes include random color perturbations
• 部分场景添加随机颜色扰动

Level 1 - Position Perturbation

Level 1 - 位置扰动

• Stove position varies more widely (broader spatial range)
• 炉台位置变化更大（空间范围更广）
• Same language expressions as Level 0
• 与 Level 0 相同的语言表达

Level 2 - Position + Language Perturbation

Level 2 - 位置+语言扰动

• Positions same as Level 1
• 位置与 Level 1 保持一致
• Language Perturbation: 6 different expressions for the same task
• 语言扰动: 使用 6 种不同的表达方式描述同一任务

 Dataset Statistics

 数据集统计

Scene Configuration

场景配置

| Difficulty Level | Number of Scenes | Position Variation | Language Perturbation | Visual Perturbation |

难度级别	场景数量	位置变化	语言扰动	视觉扰动
Level 0	6	✓ (small range)	✗	Partial (random color)
Level 0	6	✓ (小范围)	✗	部分 (随机颜色)
Level 1	6	✓ (large range)	✗	Partial (random color)
Level 1	6	✓ (大范围)	✗	部分 (随机颜色)
Level 2	6	✓ (same as L1)	✓ (6 expressions)	Partial (random color)
Level 2	6	✓ (同 L1)	✓ (6 种表达)	部分 (随机颜色)

Data Characteristics

数据特征

• Image Resolution: 256 × 256
• 图像分辨率: 256 × 256
• Camera Views:
  • 相机视图:
  • Main camera (agentview)
  • 主相机 (agentview)
  • Wrist camera (robot0_eye_in_hand)
  • 手腕相机 (robot0_eye_in_hand)
• State Dimension: 8D (6D pose + 2D gripper)
• 状态维度: 8 维 (6D 位姿 + 2D 夹爪)
• Action Dimension: 7D
• 动作维度: 7 维
• Sampling Frequency: 10 Hz
• 采样频率: 10 Hz

File Structure

文件结构

Flat-Stove-Turn-On/
├── meta/
│   ├── info.json              # Dataset metadata
│   ├── episodes.jsonl         # Episode information
│   └── tasks.jsonl            # Task descriptions
└── data/                      # Actual data (images, states, actions)
└── flat_stove_turn_on/        # bddl files

Fine-tuning with SmolVLA

SmolVLA微调

We train an "expert-only" policy, i.e., fully fine-tuning the policy head while keeping the vision encoder frozen. 训练“仅限专家”策略，即在保持视觉编码器冻结的情况下全量微调策略头。

Training Configuration

训练配置

• Model Architecture: SmolVLA
• 模型架构: SmolVLA
• Pretrained Weights: VLA-Arena/smolvla-vla-arena
• 预训练权重: VLA-Arena/smolvla-vla-arena
• Optimizer: AdamW
• 优化器: AdamW
• Batch Size: 64
• 批次大小: 64
• Learning Rate: 1e-4
• 学习率: 1e-4
• Weight Decay: $1.0 \times 10^{-10}$
• 权重衰减: $1.0 \times 10^{-10}$
• Gradient Clipping Norm: 10
• 梯度裁剪归一化: 10
• Training Steps: 40,000
• 训练步数: 40,000
• Checkpoint Saving: Every 5,000 steps
• 检查点保存: 每 5,000 步保存一次
• Warmup Steps: 1,000 steps
• 预热步数: 1,000 步

Evaluation

评估

Training Process Visualization

训练过程可视化

Training Loss Curve

训练损失曲线

The training loss drops rapidly from an initial value of 0.1, stabilizes after about 20k steps, and finally converges to around 0.005.

训练损失从初始的 0.1 快速下降，在约 20k 步后趋于稳定，最终收敛至 0.005 左右。

Learning Rate Schedule

学习率调度

The learning rate follows a warmup + decay strategy: 学习率采用 warmup + decay 策略：

• Warmup Phase: Linearly increases from 0 to 1e-4 in the first 1k steps
• Warmup 阶段: 前 1k 步从 0 线性增加到 1e-4
• Decay Phase: Afterwards, it gradually decreases to 0 following a cosine annealing schedule
• Decay 阶段: 之后按余弦退火策略逐渐减小至 0

This strategy helps the model adapt quickly in the early stage and fine-tune parameters later.

这种策略有助于模型在训练初期快速适应，后期精细调整参数。

Evaluation Setup

评估设置

task_suite_name: "flat_stove_turn_on"
task_level: 0  # can be switched to 1 or 2
num_trials_per_task: 10
init_state_selection_mode: "first"
save_video_mode: "first_success_failure"

Result Analysis

结果分析

Success Rate

成功率

| Difficulty Level | Success Rate |

难度级别	成功率
Level 0	78%
Level 1	26%
Level 2	32%

Qualitative Results

定性结果

Evaluation videos are saved in the rollout/flat_stove_turn_on/ directory, containing: 评估视频保存在 rollout/flat_stove_turn_on/ 目录下，包含：

• Success videos: marked as success
• 成功视频: 标记为 success
• Failure videos: marked as failure
• 失败视频: 标记为 failure

Success Example 成功示例

Click to watch demo video

Failure Example 失败示例

Click to watch demo video

Data Processing

数据处理

Data Processing Pipeline

数据处理流程

The overall process consists of six main stages, as detailed below:
整体流程分为六个主要阶段，具体如下：

1. Scene Definition:
1. 场景定义：
   - Create BDDL files, including region definitions, object definitions, state definitions, and image settings.
   - 创建BDDL文件，包含区域定义、对象定义、状态定义和图像设置。

2. Demonstration Data Collection:
2. 演示数据收集：
   - Control the robotic arm via keyboard in an interactive simulation environment to perform demonstrations.
   - 在交互式仿真环境中，通过键盘控制机械臂进行演示。
   - Raw trajectories are saved as HDF5 files in the `demonstration_data/` directory.
   - 原始轨迹以HDF5格式保存在demonstration_data/目录下。

3. Data Format Conversion:
3. 数据格式转换：
   - Run the `group_create_dataset.py` script to generate images via trajectory replay.
   - 运行group_create_dataset.py脚本，通过轨迹回放生成图像。
   - Image resolution is 256x256, including main camera view and wrist camera view.
   - 图像分辨率为256x256，包含相机视图和手腕视图。

4. Dataset Restructuring:
4. 数据集重构：
   - Run the `regenerate_dataset.py` script to filter out null actions.
   - 运行regenerate_dataset.py脚本，过滤掉空动作。
   - Only successful trajectories are retained, determined by whether the gripper open/close count equals 2.
   - 仅保留成功轨迹，判断依据是夹爪开合次数等于2。

5. RLDS Format Conversion:
5. RLDS格式转换：
   - Use the `tensorflow_datasets build` tool to convert the data into the standard RLDS format.
   - 使用tensorflow_datasets build工具，将数据转换为标准RLDS格式。
   - Features include: `image`, `wrist_image`, `state`/`joint_state`, `action`, `language_instruction`.
   - 特征包括：image（图像）、wrist_image（手腕图像）、state/joint_state（状态/关节状态）、action（动作）、language_instruction（语言指令）。

6. LeRobot Format Conversion:
6. LeRobot格式转换：
   - Run the `convert_data_to_lerobot_smolvla.py` script to map the data to the SmolVLA format.
   - 运行convert_data_to_lerobot_smolvla.py脚本，将数据映射到SmolVLA格式。
   - The output format contains: `observations.images.image`, `observations.images.wrist_image`, `observations.state`, `action`, `task`.
   - 输出格式包含：observations.images.image、observations.images.wrist_image、observations.state、action、task。

Each stage is sequentially connected in the main pipeline: A → B → C → D → E → F → G.
每个阶段在主流程中依次连接：A → B → C → D → E → F → G。

Common Errors

常见错误

Error message:

错误信息如下：

ValueError: Feature type 'List' not found. Available feature types: ['Value', 'ClassLabel', 'Translation', 'TranslationVariableLanguages', 'LargeList', 'Sequence', 'Array2D', 'Array3D', 'Array4D', 'Array5D', 'Audio', 'Image', 'Video', 'Pdf', 'VideoFrame']

Cause Analysis

原因分析

LeRobot v0.2.0 uses dataset feature types (such as Sequence) that conflict with the List type introduced in newer versions of datasets (≥4.0.0). Specifically, features are stored as List type, causing loading errors in newer versions.

LeRobot v0.2.0 使用的数据集特征类型（如 Sequence）与新版 datasets（≥4.0.0）中引入的 List 类型冲突。其中特征被存储为 List 类型，导致新版加载时无法识别。

Solution

解决方案

Execute the script below to modify the data types and remove the metadata from the parquet files:

执行下方脚本修改数据类型，并清除 parquet 的元信息：

Note: Adjust the features dictionary in the script according to your dataset's actual features. For example, if the dataset does not have an observation.images.image2 column, remove it. Make sure the dictionary matches the actual column names and types of the dataset.

注意：根据数据集的特征，修改脚本中的 features 字典。比如如果数据集没有 observation.images.image2 列，就移除。务必使字典与数据集的实际列名和类型一致。 Then run the script: 然后运行脚本

import os
import pyarrow.parquet as pq
from datasets import Dataset, Features, Sequence, Value, Image

root1 = "./VLA-Arena/datasets/VLA-Arena/VLA_Arena_L0_S_lerobot_smolvla/data"

# Adjust the features according to your dataset information
features = Features({
    "observation.images.image": Image(),
    "observation.images.wrist_image": Image(),          # note: this is wrist_image, not image2
    "observation.state": Sequence(Value("float32"), length=8),
    "action": Sequence(Value("float32"), length=7),
    "timestamp": Value("float32"),                       # scalar, not list
    "frame_index": Value("int64"),
    "episode_index": Value("int64"),
    "index": Value("int64"),
    "task_index": Value("int64"),
})

def fix_file(path: str):
    # read old
    table = pq.read_table(path)

    # strip metadata
    schema = table.schema.remove_metadata()
    table = table.cast(schema)

    # make HF Dataset and recast
    ds = Dataset(table).cast(features)

    # overwrite in place
    tmp_path = path + ".tmp"
    ds.to_parquet(tmp_path)
    os.replace(tmp_path, path)

    print(f"fixed {path}")

# walk through all shards
for root, _, files in os.walk(root1):
    for fname in files:
        if fname.endswith(".parquet"):
            fix_file(os.path.join(root, fname))

print("All parquet shards fixed.")

import os
import pyarrow.parquet as pq

def clean_metadata(path):
    # read table
    table = pq.read_table(path)
    # get schema without metadata
    new_schema = table.schema.remove_metadata()
    # rebuild table with new schema (data unchanged)
    new_table = table.cast(new_schema)
    # write back (without any metadata)
    pq.write_table(new_table, path, compression='snappy')
    print(f"Cleaned: {path}")

# your data directory
root = "./VLA-Arena/datasets/VLA-Arena/VLA_Arena_L0_S_lerobot_smolvla/data"

for dirpath, _, filenames in os.walk(root):
    for fname in filenames:
        if fname.endswith(".parquet"):
            clean_metadata(os.path.join(dirpath, fname))

print("All parquet files cleaned.")

License and Citation

许可证与引用

License

许可证

This dataset is built upon the VLA-Arena framework and is licensed under the Apache 2.0 License. See the LICENSE file for details.
本数据集基于 VLA-Arena 框架构建，遵循 Apache 2.0 许可证。详见 LICENSE 文件。

Citation

引用

If you use this dataset or related code, please cite the VLA-Arena project:
如果您使用本数据集或相关代码，请引用 VLA-Arena 项目：

@misc{zhang2025vlaarena,
  title={VLA-Arena: An Open-Source Framework for Benchmarking Vision-Language-Action Models},
  author={Borong Zhang and Jiahao Li and Jiachen Shen and Yishuai Cai and Yuhao Zhang and Yuanpei Chen and Juntao Dai and Jiaming Ji and Yaodong Yang},
  year={2025},
  eprint={2512.22539},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2512.22539}
}

Acknowledgements

致谢

We thank the VLA-Arena team for providing a complete framework and toolchain, making the construction of custom datasets simple and efficient.
感谢 VLA-Arena 团队提供的完整框架和工具链，使得自定义数据集的构建变得简单高效。

---

Creation Date: 2026-03-09
创建日期: 2026-03-09
Author: Qirui Bao
作者: Qirui Bao
VLA-Arena Project Page: VLA-Arena
VLA-Arena项目主页: VLA-Arena