Data Quality Control

Checking each piece of collected robot data before training to catch and discard unusable or flawed samples.

Data quality control is the step, after collection and before annotation and training, where each piece of robot data is checked and judged usable or not. Common checks include whether the camera dropped frames, whether the different sensors' timestamps are aligned, whether the joint and action recordings are complete, whether the trajectory has abnormal jumps, whether the task was actually completed, and whether the language annotation matches the video. The usual approach runs automated rule-based scripts first, followed by manual spot checks. Imitation learning will happily learn from bad demonstrations along with good ones, so quality control has a direct effect on the resulting model. In July 2026, China's Ministry of Industry and Information Technology approved industry standard YD/T 6771-2026, drafted under the lead of the China Academy of Information and Communications Technology (CAICT), which scores embodied-AI datasets across eight dimensions including completeness, consistency, and authenticity; it takes effect on November 1, 2026.

ExampleIn a teleoperated clothes-folding demonstration, the wrist camera drops frames partway through and the gripper's open/close timing no longer lines up with the video; a quality-control script flags it as unusable, so it's excluded from the training set.

Also called

Data Quality Inspection, Data Quality Review

Related

Data Cleaning · Data Curation · Valid (Usable) Data · Data Collection SOP · Multi-sensor Time Synchronization / Timestamp Alignment · Failure Data

Sources

首个具身智能数据集质量标准发布（人民邮电报，数字中国建设峰会网站转载） (Chinese)
具身智能迈向2.0：数据采集从训练场走向真实世界（科学网转澎湃新闻） (Chinese)

As of

2026-08