svla_so101_pickplace
收藏资源简介:
该数据集是使用 LeRobot 框架创建的机器人操控数据集,适用于模仿学习、机器人控制等任务。数据集包含 50 个 episode,共 11939 帧,来自 1 个任务,并包含 100 个视频(每个摄像头视角各一个视频)。数据以 30 FPS 的帧率采集,机器人类型为 so100_follower。数据特征包括:6 维动作(关节位置,如 shoulder_pan、shoulder_lift、elbow_flex、wrist_flex、wrist_roll、gripper)、6 维机器人状态(与动作维度相同)、两个摄像头图像(上方视角和侧面视角,分辨率 480x640,AV1 编码,3 通道),以及时间戳、帧索引、episode 索引、索引和任务索引等辅助字段。所有数据存储为 Parquet 文件,视频存储为 MP4 文件。数据集采用 Apache-2.0 许可证,所有训练数据未划分验证集(train 为 0:50)。
This dataset is a robot manipulation dataset created using the LeRobot framework, suitable for imitation learning, robot control, and other tasks. It contains 50 episodes, totaling 11,939 frames, from 1 task, and includes 100 videos (one video per camera view). The data is collected at a frame rate of 30 FPS, with the robot type being so100_follower. Data features include: 6-dimensional actions (joint positions: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper), 6-dimensional robot states (same dimensions as actions), two camera images (top view and side view, resolution 480x640, AV1 codec, 3 channels), as well as auxiliary fields like timestamps, frame indices, episode indices, indices, and task indices. All data is stored as Parquet files, and videos are stored as MP4 files. The dataset is licensed under Apache-2.0, and all training data is not split into validation sets (train is 0:50).
数据集概述:svla_so101_pickplace
基本信息
- 许可证:Apache 2.0
- 任务类别:机器人学(Robotics)
- 创建工具:该数据集使用 LeRobot 创建
- 数据集格式:Parquet 文件(数据存储于
data/*/*.parquet)
数据集规模
| 指标 | 数值 |
|---|---|
| 总片段数(Episodes) | 50 |
| 总帧数(Frames) | 11,939 |
| 总任务数(Tasks) | 1 |
| 总视频数(Videos) | 100 |
| 总块数(Chunks) | 1 |
| 块大小(Chunks Size) | 1000 |
| 帧率(FPS) | 30 |
数据划分
- 训练集:0:50(全部50个片段用于训练)
机器人类型
- 机器人型号:SO-100 跟随者机械臂(so100_follower)
数据特征
数据集中包含以下特征:
1. 动作(Action)
- 数据类型:float32,维度为6
- 动作名称:肩部旋转(shoulder_pan.pos)、肩部抬升(shoulder_lift.pos)、肘部弯曲(elbow_flex.pos)、腕部弯曲(wrist_flex.pos)、腕部旋转(wrist_roll.pos)、夹爪(gripper.pos)
2. 观测状态(Observation State)
- 数据类型:float32,维度为6
- 状态名称:与动作特征一致,包括肩部旋转、肩部抬升、肘部弯曲、腕部弯曲、腕部旋转和夹爪位置
3. 观测图像(Observation Images)
包含两个摄像头视角,均为视频格式:
- 上方视角(up):480×640 分辨率,3通道,AV1编码,30 FPS
- 侧方视角(side):480×640 分辨率,3通道,AV1编码,30 FPS
4. 其他元数据特征
- timestamp:时间戳(float32,1维)
- frame_index:帧索引(int64,1维)
- episode_index:片段索引(int64,1维)
- index:全局索引(int64,1维)
- task_index:任务索引(int64,1维)
文件结构
- 数据文件:存储于
data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet - 视频文件:存储于
videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4




