plug-usb-dagger-v2

该数据集是一个机器人控制数据集，使用LeRobot工具创建，专门针对双手机器人（yam_bimanual）设计。数据集包含137个训练episodes，总计33412帧，帧率为30fps。数据以parquet文件格式存储，并附带mp4格式的视频文件。数据集提供了丰富的观测和动作信息：观测包括机器人状态（32维浮点向量，涵盖左右机械臂的3D位置、6D旋转表示、夹持器状态和6个关节角度）以及来自三个摄像头的视频流（头部、左腕和右腕，每个视频为3通道RGB，分辨率360x640，H.264编码）。动作空间同样为32维浮点向量，与状态观测结构对应。此外，数据集包含时间戳、帧索引、episode索引等元数据。该数据集适用于机器人学习任务，如模仿学习或强化学习。This dataset is a robotic control dataset created using the LeRobot toolkit, specifically designed for dual-arm robots (yam_bimanual). It contains 137 training episodes, totaling 33,412 frames at a frame rate of 30 FPS. The data is stored in Parquet file format, with accompanying MP4 format video files. The dataset provides rich observation and action information: Observations include robot state (a 32-dimensional floating-point vector covering the 3D positions, 6D rotation representations of the left and right robotic arms, gripper states, and 6 joint angles), as well as video streams from three cameras (head, left wrist, and right wrist), each being 3-channel RGB video with a resolution of 360x640 and encoded with H.264. The action space is also a 32-dimensional floating-point vector, which corresponds to the structure of the state observations. Additionally, the dataset includes metadata such as timestamps, frame indices, and episode indices. This dataset is suitable for robotic learning tasks such as imitation learning and reinforcement learning.

原始信息汇总

数据集概述

基本信息

• 数据集名称: plug-usb-dagger-v2
• 许可协议: Apache-2.0
• 任务类型: 机器人学 (robotics)
• 创建工具: LeRobot (https://github.com/huggingface/lerobot)

数据集规模

• 总片段数 (episodes): 137
• 总帧数 (frames): 33,412
• 总任务数: 1
• 总视频数: 411
• 总数据块数 (chunks): 1 (每块大小 1000 个片段)
• 帧率 (fps): 30
• 数据划分: 全部 137 个片段用于训练 (train: 0:137)

机器人信息

• 机器人类型: yam_bimanual (双臂机器人)

数据结构

• 数据路径: data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet
• 视频路径: videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4

特征 (Features)

1. observation.state (观察状态)
   • 类型: float32，维度 32
   • 包含: 左右臂的位置 (x, y, z)、旋转 6D 表示 (rot6d_0 ~ rot6d_5)、夹爪状态 (gripper)、关节角度 (joint_0 ~ joint_5)
2. action (动作)
   • 类型: float32，维度 32
   • 包含: 与 observation.state 结构相同
3. observation.images.head (头部摄像头图像)
   • 类型: 视频 (h264 编码)，分辨率 360×640，3 通道，30 fps
4. observation.images.left_wrist (左腕摄像头图像)
   • 类型: 视频 (h264 编码)，分辨率 360×640，3 通道，30 fps
5. observation.images.right_wrist (右腕摄像头图像)
   • 类型: 视频 (h264 编码)，分辨率 360×640，3 通道，30 fps
6. timestamp: 时间戳 (float32，维度 1)
7. frame_index: 帧索引 (int64，维度 1)
8. episode_index: 片段索引 (int64，维度 1)
9. index: 全局索引 (int64，维度 1)
10. task_index: 任务索引 (int64，维度 1)