loop-pi05-libero_goal-task3-gated-v3-eval-cycle0
收藏资源简介:
该数据集是一个机器人学领域的数据集,专门用于机器人任务学习和控制研究。它通过LeRobot工具创建,包含20个完整的任务episodes,总计4036帧数据,帧率为20fps。数据集中记录了机器人在执行任务过程中的多模态信息:动作数据包括末端执行器的位移(dx、dy、dz)和旋转(droll、dpitch、dyaw)以及夹爪控制命令(gripper.cmd);状态观测数据包括末端执行器的位置(x、y、z)和轴角表示(axis_angle_x、y、z)以及夹爪位置(qpos_0、qpos_1);图像观测数据包括两个256x256分辨率的RGB视频流(image和image2),使用AV1编解码器。数据集还包含时间戳、帧索引、episode索引等元数据。数据以parquet文件格式存储,视频以mp4格式存储,总数据文件大小为100MB,视频文件大小为200MB。该数据集适用于机器人模仿学习、强化学习等任务,标签如libero_goal表明可能与目标导向任务相关。
This dataset is a robotics dataset specifically designed for robot task learning and control research. It is created using the LeRobot tool and contains 20 complete task episodes, totaling 4036 frames of data with a frame rate of 20fps. The dataset records multimodal information during robot task execution: action data includes end-effector displacements (dx, dy, dz) and rotations (droll, dpitch, dyaw) as well as gripper control commands (gripper.cmd); state observation data includes end-effector positions (x, y, z) and axis-angle representations (axis_angle_x, y, z) as well as gripper positions (qpos_0, qpos_1); image observation data includes two RGB video streams (image and image2) with a resolution of 256x256, using the AV1 codec. The dataset also includes metadata such as timestamps, frame indices, and episode indices. Data is stored in parquet file format, while videos are stored in mp4 format, with a total data file size of 100MB and video file size of 200MB. This dataset is suitable for tasks such as robot imitation learning and reinforcement learning, and labels like libero_goal suggest it may be related to goal-oriented tasks.
数据集概述
- 数据集名称:
JonathanGiegold/loop-pi05-libero_goal-task3-gated-v3-eval-cycle0 - 许可证: Apache-2.0
- 任务类别: 机器人学(Robotics)
- 标签: LeRobot, libero, libero_goal, pi05_libero_base
数据集结构
- 总片段数 (Episodes): 20
- 总帧数 (Frames): 4,036
- 总任务数 (Tasks): 1
- 数据分块: 每个块大小为 1,000
- 数据文件大小: 约 100 MB
- 视频文件大小: 约 200 MB
- 帧率 (FPS): 20
- 数据集划分: 训练集(片段 0 到 19)
特征说明
-
动作 (action)
- 类型:
float32 - 维度: 7
- 属性: 末端执行器位置变化 (dx, dy, dz, droll, dpitch, dyaw) 和夹爪指令 (gripper.cmd)
- 类型:
-
观测状态 (observation.state)
- 类型:
float32 - 维度: 8
- 属性: 末端执行器位置 (x, y, z)、轴角 (axis_angle_x, axis_angle_y, axis_angle_z) 及夹爪关节位置 (gripper.qpos_0, gripper.qpos_1)
- 类型:
-
观测图像 (observation.images.image 和 observation.images.image2)
- 类型: 视频
- 分辨率: 256 × 256 像素,3 通道
- 编码: AV1,像素格式 yuv420p
- 帧率: 20 FPS
- 非深度图,无音频
-
其他特征:
timestamp: 时间戳,float32类型frame_index: 帧索引,int64类型episode_index: 片段索引,int64类型index: 索引,int64类型task_index: 任务索引,int64类型
数据存储
- 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
用途与创建
- 该数据集使用 LeRobot 创建,专为机器人学习任务设计,特别是基于目标条件的机器人操作任务。



