pusht
收藏资源简介:
该数据集基于Diffusion Policy论文(Chi et al., 2024)创建,使用LeRobot框架进行数据收集与处理,采用MIT协议开源。数据集包含206个示范片段(episodes),共25650帧,帧率为10fps,所有片段属于同一任务。数据以parquet和mp4格式存储,其中视频分辨率为96x96像素,采用AV1编码。特征字段包括:视觉观测(observation.image,视频模态)、机器人状态(observation.state,包含两个电机的位置/状态)、动作(action,两个电机的控制指令)、时间戳、帧索引、完成标志、奖励、成功标志等。该数据集适用于机器人模仿学习、行为克隆、扩散策略等任务,特别是基于视觉和状态输入的机器人控制策略训练。
This dataset is created based on the Diffusion Policy paper (Chi et al., 2024), using the LeRobot framework for data collection and processing, open-sourced under the MIT license. It contains 206 demonstration episodes, totaling 25650 frames at 10fps, all belonging to the same task. Data is stored in parquet and mp4 formats, with video resolution of 96x96 pixels using AV1 codec. Feature fields include visual observation (observation.image, video modality), robot state (observation.state, containing positions/states of two motors), action (action, control commands for two motors), timestamps, frame indices, completion flags, rewards, success flags, etc. The dataset is suitable for robot imitation learning, behavior cloning, diffusion policy, and other tasks, especially for robot control policy training based on visual and state inputs.
数据集概述:live9080/pusht
- 数据集名称:pusht
- 许可证:MIT
- 任务类别:机器人学(robotics)
- 标签:LeRobot
- 首页:https://diffusion-policy.cs.columbia.edu/
- 论文:https://arxiv.org/abs/2303.04137v5
数据集结构
- 总片段数:206
- 总帧数:25,650
- 总任务数:1
- 总视频数:206
- 总块数:1(每块包含1000个片段)
- 帧率:10 FPS
- 划分:仅包含训练集(片段0至205)
数据特征
- observation.image:视频类型,形状为96×96×3(高度×宽度×通道),编码为AV1,像素格式yuv420p,无音频
- observation.state:浮点型,形状为[2],对应两个电机(motor_0、motor_1)
- action:浮点型,形状为[2],对应两个电机(motor_0、motor_1)
- episode_index、frame_index、timestamp、index、task_index:整型或浮点型,形状均为[1]
- next.reward、next.done、next.success:分别表示奖励(浮点型)、完成状态(布尔型)和成功状态(布尔型)
数据路径
- 数据文件:
data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet - 视频文件:
videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4
引用
该数据集基于 Diffusion Policy 论文,引用格式如下:
bibtex @article{chi2024diffusionpolicy, author = {Cheng Chi and Zhenjia Xu and Siyuan Feng and Eric Cousineau and Yilun Du and Benjamin Burchfiel and Russ Tedrake and Shuran Song}, title = {Diffusion Policy: Visuomotor Policy Learning via Action Diffusion}, journal = {The International Journal of Robotics Research}, year = {2024}, }




