商品上架训练数据集

内部数据采集通过手柄遥操作完成，覆盖多相机视觉流、机器人本体状态及操作指令，各传感器数据经时间戳同步后统一存储。采集流程包含任务定义、遥操作录制、数据完整性校验、标注审核三个核心环节，并实施帧率验证、逻辑一致性检查及文件完整性校验三重质量管控。数据按任务ID与版本号结构化归档，保留原始ROS2话题消息以备回溯，符合物理安全与数据管理规范，适用于多场景下的模仿学习算法训练。Internal data is collected through joystick-based teleoperation, covering multi-camera visual streams, robot body status, and operation commands. All sensor data is synchronized via timestamps before being uniformly stored. The data collection workflow includes three core stages: task definition, teleoperation recording, data integrity verification, and annotation review, and implements three-tier quality control measures including frame rate verification, logical consistency check, and file integrity verification. The data is archived in a structured manner based on task ID and version number, with the original ROS2 topic messages retained for retrospective use. It complies with physical safety and data management specifications, and is applicable to the training of imitation learning algorithms across various scenarios.

搜集汇总

背景与挑战

背景概述

该数据集是一个面向智能机器人货架补货场景的训练数据集，包含40万条通过遥操作采集的多传感器数据，覆盖视觉流、机器人状态和操作指令，每天更新。它专为零售商店或仓储超市环境设计，用于训练商品识别与陈列算法，解决条形码识别、货架适应性等问题，并采用多层安全处理确保数据脱敏和不可逆性。