U1Can weight sensing alone achieve production-grade product recognition and fraud prevention (>99% accuracy) across varying weights, shelf loads, and environmental conditions — or is sensor fusion mandatory?
🔶 PartialWeight-only is sufficient at single-kiosk scale with per-shelf calibration. Store-scale required additional signals — a sensor-fusion approach (weight + CV + session context) emerged as the minimum viable combination.
U2Can computer vision reliably attribute product interactions to specific individuals in a multi-user, open-store environment with occlusions, lighting variation, and diverse body types — using only edge inference?
⚠️ Not achievedNo. Store-level multi-user attribution did not reach the reliability threshold required for autonomous billing. Single-kiosk CV (person counting, demographics) validated successfully.
U3What is the minimum sensor density (load cells per shelf, cameras per zone) needed to achieve reliable product-level event detection without unacceptable false-positive rates?
🔶 PartialEstablished for kiosk-scale (4–6 shelves). Store-scale sensor density requirements remain open.
U4Will the thermal management system (compressor + fan + insulation) maintain safe product temperatures across the full operating range without exceeding the kiosk’s power and noise budget?
✅ ResolvedYes, validated through extended field trials with continuous temperature logging.
U5Can the system handle graceful degradation when individual components fail (sensor board offline, network drop, CV pipeline crash) without disrupting active customer sessions?
🔶 PartialPartially. Edge buffering and session recovery implemented. Full multi-component failure recovery not tested at scale.
U6What regulatory and compliance requirements (CE, EMC, electrical safety, GDPR) apply to an unattended retail IoT device — and how early must these be addressed to avoid costly late-stage redesigns?
✅ ResolvedCE/EMC testing completed at kiosk level. GDPR review conducted for camera-based systems. Full store-level compliance not pursued.
U7Can the weighing concept remain stable under fridge and freezer conditions — i.e. across sustained low temperatures, condensation, and door-cycle thermal swings?
🔶 PartialYes for refrigerated kiosks after compensation logic; freezer-grade stability required additional mechanical and firmware adaptations.
U8Is pseudo-absolute or delta-based weighing logic sufficient to avoid repeated manual recalibration during normal operation?
🔶 PartialDelta-based logic with periodic auto-zero proved sufficient under typical usage; outlier conditions still required occasional manual recalibration.
U9Are the chosen PCB architecture, ADC, microcontroller, and firmware model sufficient to meet measurement, control, and communication requirements across all kiosk variants?
✅ ResolvedThe selected platform met requirements for the validated kiosk variants; longer-term scaling to additional variants would require an architecture review.
U10Can the mechanical design achieve repeatable sensor behavior despite assembly tolerances and field handling?
🔶 PartialAchieved through fixturing, calibration jigs, and assembly procedures; remained sensitive to operator skill and incoming part quality.
U11Can the door, lock, and session mechanics be made fail-safe across reboot, brownout, network loss, and power cycling — without trapping customers or losing transactions?
🔶 PartialImplemented and validated in lab and pilot conditions; rare edge cases (simultaneous power + network loss mid-session) remained an open risk.
U12Can kiosks be remotely updated, debugged, and recovered safely at scale (including OTA firmware/software updates) without requiring on-site intervention?
🔶 PartialOTA, remote debug, and recovery flows implemented for the deployed fleet; large-scale rollout policies and rollback strategies were proven in operations but not stress-tested at hundreds of devices.
U13Is computer vision required only as an assistive signal (e.g. anti-fraud hints, demographics) or as a core recognition layer for billing?
🔶 PartialCV proved viable as an assistive signal at kiosk scale; as a core multi-user attribution layer at store scale, it did not meet the required reliability bar.