robocasa_pretrain_human300_v4_annotated5
收藏资源简介:
该数据集是一个用于机器人学领域的大规模、多模态预训练数据集,使用LeRobot工具集创建。数据集旨在支持机器人视觉-动作策略学习、模仿学习及相关模型的训练与预训练。数据内容主要来源于RoboCasa模拟环境或真实机器人平台,记录了机器人在执行各类任务过程中的交互数据。核心数据以“观测-动作”对的形式组织:观测部分包含三个固定视角(左代理视角、右代理视角、手眼视角)的RGB视频流,分辨率均为256x256,帧率为20 FPS;同时包含一个16维的浮点型机器人状态向量。动作部分对应一个12维的浮点型控制指令向量。此外,数据集中还包含了时间戳、帧索引、回合索引和任务索引等元数据,用于标识数据在时间线和任务上下文中的位置。数据集规模庞大,总计包含32,043个完整的任务执行回合(episodes),覆盖4,740种不同的任务,累计帧数超过2,900万帧。数据以Parquet文件格式分块存储,便于高效读取与处理。该数据集适用于需要大规模、多视角视觉输入与连续动作输出配对数据的机器人学习研究,特别是基于端到端学习的策略网络训练、行为克隆以及作为机器人基础模型的预训练数据源。
This is a large-scale multimodal pre-training dataset for robotics, developed using the LeRobot toolset. The dataset is designed to support the training and pre-training of robotic vision-action policy learning, imitation learning, and other related models. The data is primarily collected from the RoboCasa simulation environment or real robotic platforms, capturing interaction data generated as robots execute various tasks. The core data is organized in the form of "observation-action" pairs: the observation section includes RGB video streams from three fixed perspectives (left agent view, right agent view, eye-in-hand view), all with a resolution of 256×256 and a frame rate of 20 FPS, and also incorporates a 16-dimensional floating-point robot state vector. The action section corresponds to a 12-dimensional floating-point control command vector. Additionally, the dataset includes metadata such as timestamps, frame indices, episode indices, and task indices, which serve to identify the temporal and task context of each data entry. The dataset has a substantial scale, containing a total of 32,043 complete task execution episodes spanning 4,740 distinct tasks, with a cumulative total of over 29 million frames. The data is stored in chunked Parquet file format to enable efficient reading and processing. This dataset is suitable for robotics learning research that requires paired data of large-scale multi-view visual inputs and continuous action outputs, particularly for end-to-end policy network training, behavioral cloning, and as a pre-training data source for robotic foundation models.
数据集概述
- 数据集名称:
robocasa_pretrain_human300_v4_annotated5 - 数据集地址: https://huggingface.co/datasets/pepijn223/robocasa_pretrain_human300_v4_annotated5
- 许可协议: Apache-2.0
- 任务类别: 机器人学 (robotics)
- 标签: LeRobot, lerobot, robocasa, unified, pretrain, human
数据集结构
基本信息
- 代码版本: v3.0
- 帧率 (FPS): 20
- 机器人类型: robocasa
规模统计
- 总片段数 (episodes): 32,043
- 总帧数 (frames): 29,106,226
- 总任务数 (tasks): 4,740
- 数据分块大小 (chunks_size): 1,000
- 数据文件大小: 100 MB
- 视频文件大小: 200 MB
数据集划分
- 训练集 (train): 索引 0 至 32,043
数据特征 (Features)
观测图像 (Observation Images)
包含三组视频数据,分辨率均为 256×256 像素,3 通道 (RGB),编码格式为 H.264,帧率 20 FPS:
- robot0_agentview_left: 左侧视角
- robot0_agentview_right: 右侧视角
- robot0_eye_in_hand: 手部视角 (eye-in-hand)
动作 (Action)
- 数据类型: float32
- 维度: 12
- 帧率: 20 FPS
观测状态 (Observation State)
- 数据类型: float32
- 维度: 16
- 帧率: 20 FPS
- 状态名称: state_0 至 state_15
其他特征
- timestamp: float32,形状 [1]
- frame_index: int64,形状 [1]
- episode_index: int64,形状 [1]
- index: int64,形状 [1]
- task_index: int64,形状 [1]
数据文件路径
- 数据文件:
data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet - 视频文件:
videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4
创建工具
该数据集使用 LeRobot 框架创建。




