A Real-Robot Dataset for Multi-Stage Shoe Inspection Handling

Industrial inspection automation commonly benefits from bounded product families and controlled handling conditions. Consumer-to-consumer (C2C) resale logistics presents a complementary setting: used products can vary in geometry, packaging, and physical condition, and must be handled without damage before visual inspection. As an initial study of this setting, we present a real-robot dataset for a nine-stage shoe-inspection handling workflow, from opening a box and presenting both shoes to a camera through repacking and package transfer. The collection contains 3,507 teleoperated demonstrations totaling 67.1 hours; approximately 60 hours are used for the reported fine-tuning experiment. We fine-tune the publicly available $\pi_{0.5}$ model using its standard pipeline, without adding an explicit progress tracker, transition verifier, or recovery mechanism, as a diagnostic baseline. Across 30 trials per setting, physical workflow completion is 90.0\% on one seen shoe pair, 43.3\% on one held-out pair with aligned placement, and 26.7\% on the held-out pair without alignment. Observed failures include missed grasps, premature transitions, repeated actions, and failure to terminate. The results identify concrete physical and procedural challenges in a realistic enterprise-derived workflow while leaving defect recognition and broader product categories to future work.