fecasado/blueberry-kitchen

---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
configs:
- config_name: default
data_files: data/*/*.parquet
---

This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).

<a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path=fecasado/blueberry-kitchen">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl-dark.svg"/>
</a>

## Dataset Description

- **Homepage:** [More Information Needed]
- **Paper:** [More Information Needed]
- **License:** apache-2.0

## Dataset Structure

[meta/info.json](meta/info.json):
```json
{
"codebase_version": "v3.0",
"robot_type": "blueberry_ros",
"total_episodes": 80,
"total_frames": 29986,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 15,
"splits": {
"train": "0:80"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
"features": {
"action": {
"dtype": "float32",
"shape": [
26
],
"names": [
"l_arm_linear.x",
"l_arm_linear.y",
"l_arm_linear.z",
"l_arm_angular.x",
"l_arm_angular.y",
"l_arm_angular.z",
"l_hand_pinky",
"l_hand_ring",
"l_hand_middle",
"l_hand_index",
"l_hand_thumb1",
"l_hand_thumb2",
"r_arm_linear.x",
"r_arm_linear.y",
"r_arm_linear.z",
"r_arm_angular.x",
"r_arm_angular.y",
"r_arm_angular.z",
"r_hand_pinky",
"r_hand_ring",
"r_hand_middle",
"r_hand_index",
"r_hand_thumb1",
"r_hand_thumb2",
"base_joy.x",
"base_joy.y"
]
},
"observation.state": {
"dtype": "float32",
"shape": [
55
],
"names": [
"l_arm_j1.pos",
"l_arm_j2.pos",
"l_arm_j3.pos",
"l_arm_j4.pos",
"l_arm_j5.pos",
"l_arm_j6.pos",
"l_arm_j7.pos",
"l_hand_pinky.pos",
"l_hand_ring.pos",
"l_hand_middle.pos",
"l_hand_index.pos",
"l_hand_thumb1.pos",
"l_hand_thumb2.pos",
"r_arm_j1.pos",
"r_arm_j2.pos",
"r_arm_j3.pos",
"r_arm_j4.pos",
"r_arm_j5.pos",
"r_arm_j6.pos",
"r_arm_j7.pos",
"r_hand_pinky.pos",
"r_hand_ring.pos",
"r_hand_middle.pos",
"r_hand_index.pos",
"r_hand_thumb1.pos",
"r_hand_thumb2.pos",
"l_arm_j1.effort",
"l_arm_j2.effort",
"l_arm_j3.effort",
"l_arm_j4.effort",
"l_arm_j5.effort",
"l_arm_j6.effort",
"l_arm_j7.effort",
"l_hand_pinky.effort",
"l_hand_ring.effort",
"l_hand_middle.effort",
"l_hand_index.effort",
"l_hand_thumb1.effort",
"l_hand_thumb2.effort",
"r_arm_j1.effort",
"r_arm_j2.effort",
"r_arm_j3.effort",
"r_arm_j4.effort",
"r_arm_j5.effort",
"r_arm_j6.effort",
"r_arm_j7.effort",
"r_hand_pinky.effort",
"r_hand_ring.effort",
"r_hand_middle.effort",
"r_hand_index.effort",
"r_hand_thumb1.effort",
"r_hand_thumb2.effort",
"gaze.x",
"gaze.y",
"gaze.valid"
]
},
"observation.images.left": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 15,
"video.channels": 3,
"has_audio": false
}
},
"observation.images.right": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 15,
"video.channels": 3,
"has_audio": false
}
},
"observation.images.user": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 15,
"video.channels": 3,
"has_audio": false
}
},
"observation.images.user_gaze": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 15,
"video.channels": 3,
"has_audio": false
}
},
"timestamp": {
"dtype": "float32",
"shape": [
1
],
"names": null
},
"frame_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"episode_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"task_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
}
}
}
```

## Citation

**BibTeX:**

```bibtex
[More Information Needed]
```

This dataset contains robotic action and observation data, including arm and hand movements, gaze information, and video observations from multiple perspectives. The dataset structure shows it consists of 30 episodes with 10,911 frames, stored in parquet files and accompanied by video files. Specific data includes 26-dimensional action vectors (containing linear and angular movements of left/right arms, finger positions, etc.), 55-dimensional state observation vectors (containing joint positions and efforts, gaze data, etc.), and 480x640 resolution videos from left/right views, user view, and user gaze view.

搜集汇总

数据集介绍

构建方式

blueberry-kitchen数据集基于LeRobot框架构建，专注于机器人操作任务。该数据集包含80个完整轨迹片段，总计29,986帧数据，采样频率为15帧/秒。数据通过蓝色莓机器人平台（blueberry_ros）采集，涵盖单一操作任务。为便于处理，数据被划分为10个块（chunks），每个块容量为1000帧，并以Parquet格式存储结构化数据，同时辅以MP4格式的视频文件记录视觉信息。数据集的所有轨迹均用于训练（split为train:0:80），未设验证或测试子集，充分体现了其在单任务场景下的专注性。

特点

blueberry-kitchen数据集的显著特征在于其多模态与高维度的数据设计。动作空间为26维，涵盖双臂线性及角度运动、手指关节及底座控制指令；观测状态空间高达55维，包含双臂各关节位置与力矩、手部关节状态及用户注视信息。视觉方面，数据集提供左侧、右侧、用户视角及用户注视共四路视频流，分辨率均为640×480像素，采用AV1编解码格式，为机器人学习提供丰富的环境感知能力。此外，数据中还包含时间戳、帧索引等元信息，便于时间序列分析。

使用方法

使用blueberry-kitchen数据集时，可借助LeRobot库进行便捷加载与处理。用户通过Hugging Face平台的数据集可视化工具可预览轨迹。在应用层面，数据需按chunk索引读取Parquet文件，视频文件则按对应路径加载。该数据集适用于模仿学习、行为克隆等机器人学习场景，研究者可利用其高维动作与状态空间训练策略网络，并通过多视角图像进行视觉感知增强。由于数据集规模适中（约300MB），适合在单台机器上进行快速迭代实验。

背景与挑战

背景概述

随着机器人学习领域的迅猛发展，数据集成为推动算法进步的核心要素之一。blueberry-kitchen数据集由研究团队于近期构建，基于LeRobot框架开发，旨在为双臂协同操作任务提供高质量的演示数据。该数据集聚焦于厨房环境中的单一操作任务，包含80个演示片段，总计近3万帧，采集了机器人左右臂的关节位置、力矩、末端执行器状态以及多视角视觉信息（包括左、右相机及用户视角与注视点）。其核心研究问题在于如何利用细粒度的多模态观测数据，训练机器人完成复杂、精细的双臂操作，尤其关注动作序列的连贯性与环境交互的鲁棒性。该数据集的发布为机器人模仿学习、行为克隆及离线强化学习等领域提供了标准化基准，促进了真实场景下技能迁移研究的深入发展。

当前挑战

blueberry-kitchen数据集所解决的领域挑战在于，双臂协同操作任务中动作空间高维（26维）、观测模态异构且噪声大，传统的单臂或简化模型难以捕捉双手协调与精细操控的复杂性。构建过程中，研究者面临着多重挑战：首先，数据采集需在真实厨房环境下进行，环境光照、物体摆放及动态干扰因素难以控制，影响数据一致性；其次，通过遥操作采集高精度演示时，人手与机器人之间的动作映射存在延迟与误差，尤其在多指灵巧手和移动基座联合控制中更为突出；最后，从80个轨迹中提取通用操作策略面临样本效率低下的问题，加之多模态数据（图像、状态、力矩）在时间上的对齐与融合技术尚未成熟，给后续模型训练与泛化带来了显著困难。

常用场景

经典使用场景

在机器人学习与灵巧操作研究领域，blueberry-kitchen数据集为双臂协作与精细操作任务提供了高质量的示范数据。该数据集采集自蓝莓机器人平台，包含80个完整回合、近三万帧的机器人操作轨迹，记录了机械臂关节位置、力矩以及左右手各手指的运动指令。结合多视角视觉观测（左、右相机及用户视角），研究者可将其用于模仿学习中的行为克隆、逆强化学习等经典范式，通过端到端的方式将人类示教的动作映射为机器人的控制策略。数据集中高维的动作与状态空间设计，尤其适合探索复杂环境下多自由度机械臂的协同控制算法。

衍生相关工作

围绕blueberry-kitchen数据集，国内外研究团队已衍生出多项具有影响力的工作。在模型架构层面，部分研究者借鉴其多视角观测特性，提出了融合Transformer与卷积网络的双流视觉-运动控制模型；在数据增广方面，有工作利用数据集中的力矩信息与视觉流，开发了基于物理模型的逆动力学仿真框架，使得少量真实示教数据即可驱动生成海量训练样本。该数据集还与LeRobot生态紧密耦合，基于其统一的数据格式，催生了一批标准化评估基准，如双臂操作的成功率度量与样本效率对比协议。这些衍生工作不仅深化了对机器人示教学习本质的理解，也为后续复杂操作任务的数据集构建提供了范式参考。

数据集最近研究

最新研究方向

在机器人学习领域，blueberry-kitchen数据集为双臂协同操作与人类示教学习提供了高维度的观测与动作数据支撑。其特色在于融合了多视角视觉流（左/右/用户/用户视线摄像头）与全身关节状态（含力位混合反馈），契合当前具身智能中‘视觉-语言-动作’多模态对齐的前沿探索。伴随LeRobot框架的普及，该数据集可被用于训练基于扩散策略或Transformer架构的通用操作模型，例如模拟厨房场景中的精细抓取与双臂协调。此类研究推动了从仿真到真实世界迁移的泛化能力突破，对降低机器人数据采集成本、加速家庭服务机器人落地具有显著意义。