FiftyOne Plugins Let Teams Build Custom Robot Data Curation Tools

Robot teleoperation datasets often contain flawed episodes, such as hesitations, sensor freezes, or repeated failed grasps, that are impractical to review one by one. The article argues that generic viewers cannot flag these issues because the relevant quality checks depend on each robot's hardware and data schema. It describes the FiftyOne plugin framework, which lets developers add server-side Python tools, buttons, panels, and background jobs that run across an entire dataset. Robot data differs from image data because its structure varies widely, so plugins are presented as a way to tailor checks such as motion smoothness, sensor health, and action-state alignment. Two open-source examples, lerobot_data_curation and demo_quality_scorer, demonstrate ranking LeRobot and MCAP episodes from worst to best.