Mission model
Diameter, bends, medium, access, risk, recovery method and allowed duration. Output is a written task card, not a spoken “go look”. No card, no next act.
ROBOT · LINK · DOCK · DATA
Reconnaissance before entry, as a system
From mission setup to human review, every clip is bound to device, time, segment and software version. The MVP is not sold as “full autonomy”. Safe abort and complete evidence come first.
WHY A SYSTEM
Field failures rarely start at the propeller
Typical confined-network failures: the machine can enter but cannot return; video cannot be aligned; a gas reading has no sample location; a model flag cannot be replayed by a crew. Architecture therefore starts from the mission loop, not from a flight parameter.
The design splits an inspection into an interruptible state machine: confirm environment and abort conditions before deploy; record weak links and loss-of-link position during capture; keep raw data, AI marks and human edits afterwards. Any layer may fail. Failure must be explainable. PG-Air may later be swapped for a more mature airframe without throwing away dock procedure, data objects or asset interfaces — so the company is not bound to a single cage.
Academic and industry contrast also favours system-first. Sewer-robotics reviews typically list multi-sensor fusion, GNSS-denied localisation and digital-twin hand-off as open problems, not “fly longer”. MIT, CMU and European pipe/tunnel work discusses contact force, localisation degradation and human supervision more than consumer aerial photography. PipeGuardian uses those papers as a learning baseline. External papers and competitor manuals are not our test results.
MISSION LOOP
One mission, nine acts
Each act has its own inputs, outputs and abort right. Acts can be drilled on a rig, then chained into a full mission.
Checks and calibration
Battery, rotors, cage, sensors, clocks, link and log space. Gas sensors must carry a calibration state; cameras and thermal must record settings. Expired calibration degrades or cancels the task.
Topside deployment
Exclusion zone, safety officer, manual takeover, tether/backup and evacuation conditions. The well-head is a control surface, not a launch pad. PG-Dock’s duty starts here.
Supervised inspection
Slow advance; mark bends, obstacles and weak links. Task success outranks autonomy. First build: piloted with stability aids. “Fully autonomous exploration” is not an MVP selling point.
Abort
Loss of link, gas threshold, structural hit, heat, battery or localisation confidence can stop the run. Recovery — or marking an unrecoverable risk — is the next act, not “keep exploring”.
Recover and decontaminate
Inspect damage, quarantine anomalies, check chemistry compatibility and battery safety. Decontamination precedes data copy, so contaminants do not enter the lab or the van.
Spatial sync
Video, depth, gas, pose and segment index on one timeline. Unaligned data is reference imagery only. It does not enter a defect conclusion.
AI assist + human review
Models rank clips. Qualified people confirm. Algorithm output is not a defect finding. Edits must trace to a person, a time and a model version.
Report, work order, re-inspect
Keep raw evidence, model version, edits and before/after. Searchable in the asset system, or the mission is incomplete. Exporting a PDF is not delivery.
PRODUCT LAYERS
Four live layers plus a research line
Each layer has its own failure modes and evidence gate. Layers talk only through an ICD, so the whole vehicle is not “untestable as one lump”.
Collision, sensing, health
Replaceable cage, rotor isolation, sensing windows, quick-release battery, power cut and recovery interfaces. Health includes vibration, temperature, battery and impact counts, used to decide whether the mission continues.
Make a weak link visible
Relay / tether / cache strategy records link quality. Loss of link is not dressed up as “autonomy”. Tethered schemes such as ScoutDI show that long observation needs a physical recovery path. PipeGuardian treats that path as a switchable strategy, not default all-wireless.
Mission control and logistics
Well-head power, data offload, health check, transport, decontamination and service. Without a dock, the field degrades into “a van, a laptop, some cables” that cannot be repeated.
Evidence, assets, decisions
Segment ID, spatial index, defect workflow, model tracking, reports and customer-system interfaces. The aim is to map Shenzhen underground-pipeline data standards and Hong Kong asset-coding habits — not to invent a format no one can join.
Fire-suppressant and insulation research
Micro-extinguishing media and insulation-tape modules sit outside the MVP. Interlock, media, contamination, explosion protection, fire service, insulation and duty bounds must pass first. Until that gate, only interfaces are shown.
SENSING LOGIC
Three evidence chains, not mixed conclusions
Low light, temperature and gas are different physical quantities. The system may fuse cues. One sensor must not replace another’s conclusion.
Low-light / night vision
High-sensitivity camera, active fill and exposure control for cracks, foreign objects, sediment and surface state. Verify illuminance, motion blur, reflection and mist bounds. Raw imagery is kept; enhancement is a visualisation layer only.
Infrared thermal
Screen hot spots, temperature difference and unusual heat sources. Record emissivity, reflection, distance and ambient temperature. Thermal alone is not a gas-concentration conclusion. A thermal anomaly may trigger a gas re-check; it does not skip gas sensing.
Gas detection
Scene-selected methane, CO, H₂S, VOC or optical gas imaging. Bind calibration, response time, sample location and cross-sensitivity. Rotor downwash changes local concentration; treat it as a trial confounder, not an ignore item.
Multi-source review
Thermal, gas, video, depth and mission context form a “suspect — re-check — act” chain. Final risk judgement sits with a qualified person. That is the same logic as Hong Kong confined-space practice: permit, monitoring, rescue readiness.
APPLICATIONS
In-pipe tasks
Build evidence from controlled scenes, then extend to harder media. This is not façade, overhead-line or open-air inspection.
Large pipe tunnels / utility corridors
Pre-entry reconnaissance and re-inspect for obstacles, cracks, seepage, heat rise and weak links. Cables, water and fire mains share a corridor. Asset coding outranks a pretty 3D model.
Underground water / drainage
Prefer isolated, dry, non-explosive segments. Sewage and wet environments re-define decontamination, sealing and recovery. Biofilm, flow and gas stratification change sensing validity.
Industrial passages / tunnel service ducts
Release case by case against site HSE, medium, gas and electrical classification. “A drone can fit” does not replace a work permit. Factory permit systems are often stricter than municipal networks. That is a feature: it forces abort conditions into the product.
MVP · IN SCOPE
First verifiable envelope
- Dry, de-energised, non-explosive;
- Target ID ≥600 mm (unvalidated);
- Stable structure, clear access and recovery;
- 30 m straight pipe, bends, controlled obstacles;
- Human supervision and pre-set abort rules.
NOT YET · OUT OF SCOPE
Not in the first promise
- Pressurised gas and explosive atmospheres;
- Fully submerged or strongly corrosive media;
- Long unattended runs or in-pipe battery swap;
- Unvalidated centimetre accuracy or continuous coverage;
- PG-X live fire-suppression or insulation intervention.
INTERFACES
Swap the airframe, keep the job
If later tests show another platform fits a diameter or medium better, PipeGuardian still keeps the task card, evidence object and well-head procedure. That is the line between a systems company and a model-aircraft shop.
The ICD at least covers: mechanical quick-release and CG budget, power and e-stop, time sync, telemetry dictionary, payload safety interlock, data-offload format. PG-Atlas keys on the segment, not on a flight log. A later crawl or tethered fill-in can still merge into the same evidence object.
More sensing, flight-control, comms and recovery detail sits on technology. Permits, gas and loss of link sit on safety bounds. Target numbers sit on target specifications.
Standards: Design and pilots map GB 46768-2025, GB 30871-2022, GB 12358-2024, GB/T 3836.1-2021, GB 3836.15-2024, GB/T 4208-2017, DB4403/T 587-2025 and the Hong Kong confined-space code of practice by scene. Applicability is decided by the site owner and qualified persons. “Referenced” is not “certified”.
Relation to the spec page: this page names objects and failure modes; that page lists unverified numbers. Any edit that wants to write endurance or accuracy in the present tense here should be sent back to the evidence gate. The concept-model image in this narrative only explains module placement. It remains not a production photograph.