axiboai/piper-stack-act-tactile-v0

该数据集是一个机器人操作数据集，包含400个训练片段，总计223,894帧数据，采样频率为50Hz。数据集包含多模态观测数据：来自左腕、右腕和顶部场景摄像头的224x224 RGB视频（H.264编码）、7维机器人状态（关节位置和夹爪状态）、32维触觉传感器数据，以及7维动作指令（关节目标位置）。数据集还包含丰富的元数据：演示者ID、工作站ID、任务变体、毛巾ID、光照条件、垫子ID、任务成功标志和注释。数据以Parquet格式存储，视频以MP4格式存储。This is a robotic manipulation dataset containing 400 training episodes, with a total of 223,894 frames sampled at 50Hz. The dataset includes multimodal observational data: 224×224 RGB videos (H.264 encoded) captured by left wrist, right wrist, and top-down scene cameras, 7-dimensional robot state (joint positions and gripper status), 32-dimensional tactile sensor data, and 7-dimensional action commands (joint target positions). Additionally, the dataset provides comprehensive metadata including demonstrator ID, workstation ID, task variant, towel ID, lighting conditions, mat ID, task success flag, and annotations. Structured data is stored in Parquet format, while video files are stored in MP4 format.