simple_grasp_224
收藏资源简介:
该数据集是使用LeRobot框架创建的机器人学数据集,旨在支持机器人控制与操作任务的研究与开发。数据来自名为stringman的机器人执行任务时收集的交互数据,包含370个完整情节(episodes),共计148,916帧,覆盖90个不同任务。数据以分块形式组织,每块约1000帧,总数据文件大小约100 MB,视频文件大小约200 MB,采集帧率为30 FPS。所有数据均划分为训练集。核心内容包括机器人动作与观测状态序列:动作是一个5维浮点向量(表示x、y、z方向速度、手腕速度和手指速度);观测状态是一个43维浮点向量(包含机器人本体状态如速度、位置、旋转、角度、压力、张力等,以及多个目标物体的方位和距离信息);观测图像来自夹爪摄像头的视频数据,每帧为3通道RGB图像,分辨率224x224,视频编码AV1,像素格式yuv420p,无音频;时间与索引信息包括时间戳、帧索引、情节索引、全局索引和任务索引。数据以Parquet格式存储,视频以MP4格式存储。适用于机器人模仿学习、强化学习、视觉运动控制等任务,特别适合研究多模态感知下的抓取与操作策略。数据集采用Apache-2.0许可证。
This dataset is a robotics dataset created using the LeRobot framework, aimed at supporting research and development in robot control and manipulation tasks. The data consists of interaction data collected from a robot named stringman during task execution, containing 370 complete episodes, totaling 148,916 frames, and covering 90 different tasks. The data is organized in chunks, each with approximately 1000 frames, with a total data file size of about 100 MB and video file size of about 200 MB, captured at a frame rate of 30 FPS. All data is divided into a training set. The core content includes sequences of robot actions and observation states: the action is a 5-dimensional floating-point vector (representing velocities in x, y, z directions, wrist speed, and finger speed); the observation state is a 43-dimensional floating-point vector (including robot body states such as velocity, position, rotation, angle, pressure, tension, etc., as well as orientation and distance information for multiple target objects); observation images come from gripper camera video data, each frame being a 3-channel RGB image with a resolution of 224x224, video encoded in AV1, pixel format yuv420p, without audio; time and index information includes timestamp, frame index, episode index, global index, and task index. The data is stored in Parquet format, and videos are stored in MP4 format. It is suitable for tasks such as robot imitation learning, reinforcement learning, and visual-motor control, particularly for studying grasping and manipulation strategies under multimodal (state + visual) perception. The dataset uses the Apache-2.0 license.
数据集概述:naavox/simple_grasp_224
- 许可证:Apache-2.0
- 任务类别:机器人学 (Robotics)
- 标签:LeRobot
- 创建方式:使用 LeRobot 框架创建
数据集详情
| 属性 | 值 |
|---|---|
| 总片段数 (Episodes) | 370 |
| 总帧数 (Frames) | 148,916 |
| 总任务数 (Tasks) | 90 |
| 帧率 (FPS) | 30 |
| 机器人类型 | stringman |
| 训练集划分 | 片段 0 至 369 (全部) |
| 数据文件大小 | 100 MB |
| 视频文件大小 | 200 MB |
| 数据格式 | Parquet (分块) |
| 视频格式 | MP4 (AV1 编码, 224x224 分辨率) |
特征维度说明
- 动作 (Action):5 维,浮点型,包含线速度 (vel_x/y/z)、腕部速度 (wrist_speed)、手指速度 (finger_speed)。
- 观测状态 (Observation.State):43 维,浮点型,包含速度、夹爪位置与旋转、手指角度、激光测距、压力、腕部角度、目标力、多个物体(如篮子、玩具箱、垃圾桶、游戏手柄、停车位置)的方位与距离、摆动取消标志、张力值 (4 个)、龙门架位置 (3 轴)、视觉位置 (3 轴)、悬挂位置 (3 轴)。
- 观测图像 (Observation.Images.gripper_camera):视频类型,3 通道,224×224 分辨率,帧率 30 FPS,AV1 编码。
- 辅助字段:时间戳 (timestamp)、帧索引 (frame_index)、片段索引 (episode_index)、全局索引 (index)、任务索引 (task_index),均为 1 维。
数据文件结构
- 数据路径:
data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet - 视频路径:
videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4 - 块大小:每个块包含 1000 帧
参考文献与更多信息
- 主页:暂无
- 论文:暂无
- 引文格式:暂无 (BibTeX 信息待补充)




