axiboai/piper-stack-scripted-smolvla-v3

该数据集是使用LeRobot创建的机器人操作数据集，专门用于机器人学习任务。数据集包含200个训练片段，总计125,272帧数据，帧率为20fps。数据特征包括：来自四个摄像头的视频观测（左腕、右腕、前场景、顶场景，分辨率224x224，3通道，H.264编码），机器人状态观测（30维，包括关节位置、速度和夹爪力），末端执行器位姿（14维），动作数据（关节位置目标14维和末端增量12维），以及丰富的元数据（如演示者ID、工作站ID、任务变体、毛巾ID、光照条件、垫子ID、成功标志和注释）。数据集以Parquet格式存储，适用于机器人模仿学习或强化学习研究。

This dataset is a robotics manipulation dataset created using LeRobot, designed for robot learning tasks. It contains 200 training episodes with a total of 125,272 frames at 20fps. Features include: video observations from four cameras (wrist left, wrist right, scene front, scene top, resolution 224x224, 3 channels, H.264 encoded), robot state observations (30-dimensional, covering joint positions, velocities, and gripper forces), end-effector poses (14-dimensional), action data (joint position targets 14-dimensional and end-effector deltas 12-dimensional), and extensive metadata (such as demonstrator ID, station ID, task variant, towel ID, lighting conditions, mat ID, success flag, and notes). The dataset is stored in Parquet format and is suitable for imitation learning or reinforcement learning research in robotics.