SENSESensing: three physical quantities, separately
Visible / night vision answers “what is on the surface”. Thermal answers “where is the temperature difference”. Gas sensors answer “is the target gas present”. Depth / LiDAR answers geometry. Fusion only supplies context. Low-light enhancement must not overwrite the raw image. The gas sample port must account for rotor wash and the wall boundary layer. Mixing those quantities into one present-tense detection-rate sentence conflicts with the evidence page.
CONTROLFlight control: stability below the job, abort above autonomy
First edition: human supervision plus assistive stability. Attitude, height / pipe-axis hold and wall-follow are separate modes, not a default full-autonomy explore. Localisation confidence drop, gas threshold, low battery and loss of link can all interrupt. GNSS is unavailable in pipe. Localisation must carry a degradation alarm. The 0.2–3 m wall-hold figure is a control-mode target on the specs page. It is not a verified centimetre claim.
LINKCommunications: a weak link must be visible
Underground bends and standing water can collapse radio quality. PG-RELAY starts with tethered / wired backup and a few recoverable nodes, records RSSI / loss / delay, and marks last-heard position. This site does not promise a full underground mesh. Tethered products such as ScoutDI show that a physical recovery path is part of the comms architecture. That architecture contrast is a classroom, not a PipeGuardian field result.
RECOVERRecovery: a design input, not after-sales
A grabbable cage, orientation fool-proofing, well-head guidance, emergency power cut, battery isolation after impact. Kit that cannot be recovered is a new hazard in a confined space. The dock handles power, unload, inspection and decontamination so a job can repeat, instead of improvising cables each time. Recovery belongs on the safety bounds page as a permit input, not a brochure afterthought.
SYNCData sync: unaligned data is not evidence
Video, point cloud, gas, attitude, link quality and operator input must share a timeline and a segment index. PG-Atlas keeps model version, confidence, human edits and raw-file hashes. Failed alignment stays a reference. It does not enter a defect conclusion. Atlas aims to map Shenzhen underground-line coding, not a private format that cannot join an owner ledger.