守Layer 6 The Autonomy Stack · self-maintaining

Care. Make the building the robot's map and fixture, swap whole organs, and keep a remote human for exceptions.

8 subsystems28 projects29 papers31 numbers70 sources

L05 · concept visual · AI

守01 Thesis

A fully automated habitat does not mean a humanoid fixing everything; it means few, captive robots doing scheduled tasks on fixtures the building provides, organ-level swaps instead of repairs, a fleet ledger that learns real failure rates, and a remote operator for the few percent of exceptions. The most scaled autonomous maintenance robots today (row-captive waterless PV cleaners with billions of cleanings) succeed because they are mechanically constrained, not because they are smart. Space shows the same pattern: Canadarm2/Dextre work because ORUs were designed for robots, and Astrobee needed crew for only about 20% of its console time. The non-obvious insight is that design for robot maintenance is mostly the building's job: datum sockets, rails, one handle standard, and coarse ORU granularity turn hard perception and manipulation into easy positioning. General-purpose home manipulators (NEO, Helix, LBMs) are improving fast, but in 2026 they still run on teleoperation, so ORIJO should treat them as an upgrade path, not a dependency.

3.89 billion panel cleanings

Ecoppia row-captive waterless PV robots, cumulative panels cleaned

claimedecoppia-site

about 80 %

Astrobee console time without crew (>1,200 console h vs ~220 crew h)

calculatedastrobee-tfr

18 %

ISS ECLSS ORUs that failed more often than predicted (6% by >4x)

measuredowens-iss

Slab 6 of the stack · see all eight

Maturity of options assessed

commercial 11
pilot 13
experimental 4
speculative 0

Deep pageTSURU ridge robot · walking in 3D

守02 Subsystem explorer

How the care layer works, option by option.

Each subsystem lists the technologies the research found, with maturity, an estimated technology readiness level and the numbers that matter. Hover a TRL bar for the reason behind it.

CARE-S1 · function

Sense and patrol (inspection)

Detect degradation (leaks, hot joints, corrosion, cracks, loose latches, soiling) before it becomes failure, on a schedule and on demand.

How it works

A sensor head is carried past every asset on a repeatable route, and each reading is registered to the asset's ID so trends, not snapshots, drive work orders. Thermal imagers read emitted long-wave IR to find resistive hot spots and blocked flow; ultrasonic microphones (to hundreds of kHz) pick up gas leaks and partial discharge that are inaudible; contact ultrasonic transducers measure wall thickness by time-of-flight on steel. The carrier choice is a trade between access and repeatability: fixed sensors are cheapest per reading but blind between points; legged robots climb stairs built for humans; magnetic crawlers hold contact for UT on steel; drones see envelopes and roofs. In ORIJO the building's own crease rails and 910 mm datum sockets make the route mechanically repeatable, so a rail-captive head returns to within the socket tolerance every time and image differencing becomes trivial.

COMMERCIALTRL 9

Legged inspector (ANYmal, Spot)

Quadruped carrying thermal, zoom, ultrasonic-acoustic and gas payloads on autonomous routes with dock recharge

90 (90–120 min walking, 2 km) min reportedANYmal runtimeanymal-spec
100 min reportedANYmal 70% quick chargeanymal-spec
0–384 kHz reportedANYmal ultrasonic mic rangeanymal-spec
90 / 60 min reportedSpot runtime / rechargespot-spec
14 kg reportedSpot payloadspot-spec

Best for E7 steppe/farm clustersE3 desert sites with many moduleshotel clusters (rented per visit)

Limits Runtime 90 min sets a duty cycle of roughly 1:1 run:charge for Spot; IP54 (Spot) is marginal for driving rain and snow; capital cost too high to dedicate to one home.

Sold and operated on offshore, energy and manufacturing sites; >1,500 Spot units in customer hands (claimed).

EXPERIMENTALTRL 3

Rail-captive sensor head (TSURU / crease rail)

Robot locks to datum sockets on the ridge rails; the structure is the map

25–35 kg targetTSURU target massv3-concept
60 W targetTSURU target avg powerv3-concept

Best for all ORIJO environments with YOSHI ridgesE4 polar and E5 mountain where walking outside is unsafe

Limits Sees only what is ridge-adjacent; one robot per contiguous rail network; untested in icing.

Concept from v3 M09 with analogues (ARMADAS lattice robots, BMU gantries, Ecoppia row robots) but no ORIJO hardware built.

COMMERCIALTRL 9

Magnetic / climbing crawler (Gecko TOKA class)

Magnetic wheels hold sensor arrays (UT, EMAT, visual) against steel walls, tanks, boiler tubes

125 USD million reportedGecko Series Dgecko-trr
1.25 USD billion reportedGecko valuationgecko-trr

Best for steel tanks, ASHI steel legs, battery cabinet enclosures (rare in ORIJO)

Limits Needs ferrous substrate; ORIJO panels are LVL/composite, so relevance is limited to steel legs and tanks.

Gecko serves US Navy, power and refinery customers at unicorn scale; $300M Navy IDIQ (claimed).

COMMERCIALTRL 8

Inspection drone

Photogrammetry and IR thermography of envelopes and roofs

269 papers measuredRobotic built-environment inspection papers reviewedhalder

Best for E1 rooftopE7 farmland clusterspost-storm envelope surveys

Limits Wind, snow, rules on unattended flight; sees surfaces only, cannot touch or fix.

Widely used for building energy audits (Rakha 2018 review); autonomous dock-in-a-box products exist but regulation limits unattended BVLOS flight.

PILOTTRL 7

Intra-vehicular free flyer (Astrobee)

Fan-propelled cube robots that dock to recharge and survey a habitat interior

up to 2 h reportedAstrobee flying timeastrobee-tfr

Best for E8 lunar / TSUKI interiors

Limits Microgravity only; localization in a changing interior was its biggest continuing challenge.

Operating on ISS since 2019 for research; not an operational caretaker yet.

CARE-S2 · function

Surface keeping (PV, glazing, facade)

Keep energy and daylight surfaces clean and keep envelopes inspected without people on ladders or ropes.

How it works

Soiling is a deposition process whose rate depends on dust load, humidity (cementation after dew) and tilt; losses grow roughly linearly until rain or cleaning resets them, so the optimum is frequent light cleaning when water is scarce. The most mature autonomous answer is the row-captive robot: it rides rails on the PV row, sweeps a rotating microfibre cylinder whose helical fins create a directional airflow that pushes dust off the low edge, and parks in a dock at the row end each morning. Because it is captive to the row, it needs no perception beyond end-stops and gap sensors, which is why it reached billions of cleanings. Facades are the opposite case: a building maintenance unit (BMU) gantry positions an industrial arm with force-torque control against glass, with a human supervising from the roof.

COMMERCIALTRL 9

Row-captive waterless PV robot (Ecoppia class)

Microfibre + airflow cylinder on a row-mounted frame, nightly, docked

3890000000 panels claimedPanels cleaned (cumulative)ecoppia-site
10510000 sessions claimedAutonomous cleaning sessionsecoppia-site
~100 ft2/min (about 9.3 m²/min) reportedCoverage rate per robot (Ketura 2014)ecoppia-ketura
0 L reportedWater usedecoppia-ketura
20–30 % claimedHelix vs straight fin motor power savingpat-evermore

Best for E3 desertE7 steppeMIURA PV wings and BYOBU fences laid out as straight rows

Limits Needs straight continuous rows with a rail edge; module vibration from cleaning has been studied as a microcrack risk (Figgis 2023, not verified in detail).

Fleet counters claim 3.89 billion panel cleanings across 5 GW+; operating since 2014.

PILOTTRL 7

BMU-mounted robotic arm (Skyline Ozmo)

Industrial arm on a gondola with lidar, ultrasonic, force-torque sensing (sill sensors unverified); human supervisor on roof

~3x ratio claimedSpeed vs traditional cleaningozmo-trr

Best for E1 rooftop towers (not low ORIJO modules)

Limits Needs a BMU; lidar struggles on transparent glass; still supervised.

Operating on one 45-storey Manhattan tower; claims ~3x faster than manual.

COMMERCIALTRL 9

Brush crawler / hand-portable PV robot

Battery crawler with rotating brush moved row to row by a person

Best for irregular arrays, slopes, KUMO racks

Limits Human carries it between rows; not autonomous between arrays (specs unverified in this sweep).

Several vendors sell them (e.g. SolarCleano); specs not verified here.

PILOTTRL 6

Fixed-gantry facade robot (Fraunhofer SIRIUSc)

Fully automated roof gantry secures, powers and positions a facade-walking robot

4000 m² reportedFraunhofer HQ facade areasirius

Best for reference for gantry + robot pairing

Limits Bespoke to one building; cost not reported.

Demonstrated on an 80 m, 4,000 m² facade (2008); no evidence of commercial scale.

CARE-S3 · function

Swap-in maintenance (organ / ORU replacement)

Repair by replacement: remove a failed unit as a whole, install a spare, repair the old unit at a depot.

How it works

Orbital replacement units (ORUs) concentrate failure-prone parts into boxes with robot-compatible grapple fixtures, alignment guides and captive fasteners, so a manipulator with a tool changer (Dextre) can do what used to need a spacewalk. The physics is constraint: a coarse cone or V-guide removes position error mechanically, then a fine kinematic seat fixes all six degrees of freedom, then a single driven fastener preloads it. Connectors must be blind-mate and dry-break so fluids and power survive misalignment. NASA's Astrobee team found the trap: making many small components into ORUs added captive fasteners, access geometry, mass and size, and in retrospect they would have made only a few larger subassemblies replaceable. ORIJO's JO-PORT already sets the right granularity: the organ (water unit, battery cabinet, farm cassette) is the ORU.

COMMERCIALTRL 9

Manipulator with tool changer on grapple fixtures (Canadarm2 + Dextre)

Relocatable 7-joint arm walks end over end between power-data grapple fixtures; dexterous two-arm robot swaps ORUs, mostly ground-commanded

17.6 m reportedCanadarm2 lengthmss-wiki
7 joints reportedCanadarm2 jointsmss-wiki
116000 kg reportedCanadarm2 max payloadmss-wiki

Best for E8 lunarreference for TSURU end-over-end walking

Limits Space-qualified cost; slow, scripted, ground-supervised operations.

Operational on ISS since 2001 (arm) and 2008 (Dextre). Note: an operational agency system, not a commercially sold product; 'COMMERCIAL' here means flight-proven in routine operations.

EXPERIMENTALTRL 3

Organ-level JO-PORT swap (v3 M04)

Quasi-kinematic ball/V-block datum, floating port box, dry-break couplers; robot or two people roll organ in

± 3 mm targetPort box floatv3-concept

Best for all ORIJO typologies

Limits Organs up to hundreds of kg exceed TSURU payload; needs cart or crane for heavy organs.

All parts commercial (ISO 16028 couplers, kinematic mounts); integrated socket untested.

PILOTTRL 7

Component-level ORUs everywhere

Every board, fan and sensor is field-replaceable

Best for avoid except for wear parts

Limits Mass, captive fasteners and access geometry grow; interfaces you forget (EMI, acoustics) break modularity.

Astrobee flew it; its own team judged it over-applied.

CARE-S4 · function

Mobile manipulation and housekeeping

Do the open-ended indoor physical tasks: tidy, carry, wipe, load, fetch, reset a room.

How it works

Three hardware classes exist: wheeled carts that only carry (Relay, Servi), wheeled single-arm manipulators (Stretch), and humanoids (NEO, Figure 03, Digit). The 2023–2026 change is in the policy: diffusion policies generate action chunks by iteratively denoising from noise conditioned on camera images, which handles multimodal demonstrations; large behaviour models pretrain one diffusion transformer on hundreds to thousands of hours of teleoperated demos, then finetune a new task with a fraction of the data; VLA models (Helix, pi0.5) split a slow vision-language planner (7–9 Hz) from a fast visuomotor controller (200 Hz). Reliability is still the gap: real deployments run with a human teleoperator who takes over unfamiliar tasks, and every takeover becomes training data.

COMMERCIALTRL 9

Delivery cart robot (Relay, Bear Servi)

Wheeled AMR with lidar navigation, lockable bin, elevator integration

51 % reportedLG stake in Bear Roboticsbear-rave

Best for large hotels; not a 6-key ORIJO hotel

Limits Carries only; value scales with corridor length, which ORIJO lacks.

Deployed in hotels and restaurants, plus 10 hospitals at Relay's 2022 launch; LG took 51% of Bear Robotics (Jan 2025).

PILOTTRL 6

Wheeled single-arm manipulator (Hello Robot Stretch 3)

Telescoping arm on a lift mast, two-wheel base, dexterous wrist

24950 USD reportedStretch 3 pricestretch-spectrum
24.5 kg reportedStretch 3 massstretch-spectrum

Best for KURA core and farm bays with level floors

Limits No stairs; low payload; needs level thresholds (fits ORIJO's flat jo floor).

Sold for research; vendor says consumer homes are not its market yet.

PILOTTRL 6

Home humanoid with teleoperation fallback (1X NEO)

Tendon-driven humanoid; remote 'Expert Mode' operator guides unknown tasks

20,000 or 499/month USD claimedNEO priceneo-newatlas neo-kalil
30 kg claimedNEO massneo-newatlas
~4 h claimedNEO batteryneo-newatlas
22 dB claimedNEO noiseneo-newatlas

Best for future home and hotel back-of-house

Limits Autonomy share undisclosed; privacy of remote viewing; unit economics of teleoperators.

Preorders 2025, US deliveries planned 2026; WSJ found it teleoperated for most of the visit.

PILOTTRL 7

Industrial humanoid (Agility Digit)

Bipedal tote handler with independent safety controller and fast-charge dock

9 min claimedDigit 5 charge timedigit5
10:1 ratio claimedDigit 5 run:chargedigit5
>100,000 at ~98% accuracy totes claimedDigit 4 totes at GXOdigit5

Best for farm logistics (crate moving) in larger HATA-MURA sites

Limits Warehouse-structured tasks only.

>65,000 operating hours at customer sites (claimed) in warehouses.

EXPERIMENTALTRL 5

Learned policies (Diffusion Policy, TRI LBM, Helix, pi0.5)

Imitation learning from teleop demos with diffusion/flow heads and VLM planners

46.9 % measuredDiffusion Policy average improvement over prior SOTAdp
<30 % of task data calculatedLBM finetune data to match single-task (sim, unseen tasks)lbm
1695 h of demos reportedLBM pretraining mixturelbm
~500 h reportedHelix teleop training datafigure-helix

Best for fixed-geometry repeated tasks: cartridge swaps at identical JO-PORTs

Limits Success rates vary by task; distribution shift still degrades; evaluation is expensive (~1,800 real trials in the LBM study).

Strong lab and pilot results; no published long-run home reliability data.

CARE-S5 · function

Prognostics: digital twin and failure ledger

Turn telemetry and inspection data into predicted failures, spare-part counts and work orders.

How it works

Rule-based fault detection (e.g. APAR rules for air handling units) flags known fault signatures; machine-learning condition models estimate remaining useful life; a planner schedules the swap before failure. The spares side is Bayesian: each unit type starts with a prior failure rate and an error factor; every hour without failure and every failure updates the posterior, as NASA's ISS logistics team does for ECLSS ORUs. Fleet data matters more than per-site intelligence: ISS experience moved estimates both ways (26% of ORUs proved twice as reliable, 18% worse than predicted, one 7.8x worse) and the corrected rates cut Mars-class spares mass by 3.9–6.0 t. ORIJO's advantage is identical organs across many sites, so the ledger converges quickly.

PILOTTRL 6

BIM/IoT digital twin with AFDD (Hosamo 2022)

APAR rule FDD + ML condition prediction + maintenance planning on BIM

26 months measuredField data spanhosamo

Best for KURA HVAC/HRV, heat pump

Limits BIM integration effort; site-specific tuning.

Tested on a real Norwegian building with Aug 2019-Oct 2021 data.

COMMERCIALTRL 9

Bayesian fleet failure-rate ledger (ISS MADS approach)

Prior + operating hours + failures → posterior failure rate → spares and swap interval

18 % measuredISS ECLSS ORUs worse than predictedowens-iss
6 % measuredORUs >4x worse than predictedowens-iss
26 % measuredORUs at least 2x more reliable than predictedowens-iss
3.9–6.0 t calculatedSpares mass saved, 1,200-day Mars missionowens-iss

Best for every ORIJO organ, all environments, especially remote E4/E5

Limits Needs honest failure logging and enough identical units.

Used operationally by ISS logistics (MADS); the method is mature but has no off-the-shelf product for building fleets, so ORIJO's DAICHO implementation is TRL ~4.

PILOTTRL 7

Acoustic and thermal anomaly mapping

Beam-forming microphones or ultrasonic mics locate leaks and failing mechanisms; IR finds hot joints

Best for hinge and fan bearing wear, water leaks, DC connector heating

Limits Needs baselines per asset; noisy environments.

ANYmal ships acoustic imaging option; SoundSee flew on Astrobee.

CARE-S6 · function

Remote operations and teleoperation fallback

Provide a human in the loop only for exceptions, from a centre that serves many sites.

How it works

Every autonomous task has a confidence gate; below it the robot pauses in a safe state and pages an operator, who either issues a goal-level command ('go to this pose', as Astrobee operators do because of latency) or takes direct teleoperation. The operator's attention is the scarce resource, so a scheduler ranks which robot needs help most (scaled autonomy). Each intervention is logged as a demonstration and fed back into policy training, which is how 1X and Figure gather data. Astrobee shows the target shape: over 1,200 h of console support with only ~220 h needing crew, i.e. nearly 80% of operations needed no person at the site.

PILOTTRL 8

Ground-operator model (Astrobee/ISS)

Plans run autonomously; ground teleoperates simple tasks and recovery

>165 operations measuredAstrobee real-time operationsastrobee-tfr
>1,200 / ~220 h measuredConsole hours / crew hoursastrobee-tfr
~80 % calculatedShare of console time without crewastrobee-tfr

Best for all ORIJO remote sites via Starlink-class links

Limits Latency forces goal-level commands, not joystick control.

Routine on ISS since 2019 but as a NASA research facility, not a commercial service; ~20% of console hours still needed crew. Terrestrial multi-site remote ops for homes is a transfer, not the same system.

PILOTTRL 6

Consumer teleop fallback (1X Expert Mode)

Company operator sees through robot cameras and guides tasks

Best for home and hotel interior tasks

Limits Privacy, trust (WSJ: 'noticeably teleoperated for most of her visit'), labour cost per robot.

Announced 2025; operating share undisclosed.

PILOTTRL 7

Remote supervision in assisted living (MoveCare field study)

Remote supervisors monitor robots in private homes

130 weeks measuredField study lengthluperto
10 homes measuredApartmentsluperto

Best for homes with elderly residents

Limits Supervision raises acceptance but costs staff time.

130-week, 10-apartment deployment.

CARE-S7 · function

Passive durability and self-healing materials

Reduce the number of maintenance tasks to begin with.

How it works

Bacteria-based concrete carries alkali-tolerant spores (e.g. Bacillus) plus a calcium-lactate nutrient; when a crack lets water in, spores germinate and metabolise lactate into CO₂, which precipitates as calcium carbonate and seals the crack. Intrinsic self-healing polymers reform reversible bonds (Diels-Alder, hydrogen bonds, disulfides) when warmed or pressed; extrinsic ones release healing agent from microcapsules. These heal cracks and scratches, not structural failure, and capacity is limited to a few cycles or sub-millimetre gaps. In ORIJO, concrete appears only in ASHI ground pads and some ballast; self-healing elastomers are more relevant to hinge membranes and coupler seals.

COMMERCIALTRL 8

Bacteria-based self-healing concrete (Basilisk)

Spores + calcium lactate precipitate CaCO3 in water-exposed cracks

up to 1 mm claimedCrack width healedbasilisk
15 years claimedBus-lane life extensionbasilisk

Best for ASHI pads, KUMO water tanks, E7 farm slabs

Limits Needs water to activate; cost premium not stated; heals non-structural cracks.

Sold as admixture, mortar and liquid repair; field uses in a wastewater settling tank and a bus lane.

PILOTTRL 5

Self-healing polymers and coatings

Reversible covalent/supramolecular bonds or microcapsules

Best for KASANE hinge membranes, gasket faces

Limits Healing often needs heat or contact pressure; UV and cold resistance unproven for architecture.

Coatings exist commercially in niches; structural use remains research (Goyal 2022 review).

COMMERCIALTRL 9

Sacrificial wear parts by design

Put wear into cheap swappable parts (wipers, seals, brushes, filters) sized to a fixed interval

Best for every ORIJO organ

Limits Requires logistics of consumables.

Standard industrial practice.

史03 Projects · including failures

Who has tried this, and what happened.

28 projects and companies, oldest first. 1 of them closed, failed or went bankrupt; they are kept in, because investors learn more from them than from the successes.

When	Project	What	Status	Outcome or lesson
2000–2026	ISS ECLSS logistics and maintenance (MADS)NASA · US	Bayesian failure-rate updating for ORUs	active	Operations revealed underestimated failure rates (one ORU 7.8x) and saved 3.9–6.0 t spares for Mars-class missions. owens-iss
2001–2026	Canadarm2 + Dextre (Mobile Servicing System)CSA / MDA / NASA · CA	Relocatable arm and two-armed dexterous robot that replace ORUs outside the ISS	active	Most robotic ops are ground-commanded; ORU design (grapple fixtures, standard bolts) is what made robot maintenance possible. mss-wiki yt-dextre oos-ma
2002–2026	RoombaiRobot · US	Consumer floor-cleaning robots	restructured (bankruptcy; acquired by contract manufacturer)	Filed Chapter 11 on 2025–12–14; emerged by Jan 2026 as a privately held, wholly owned subsidiary of its contract manufacturer and lender Shenzhen Picea, still selling and supporting Roombas. The independent company failed; the product survived as a commodity. Lesson: a one-task home robot is a manufacturing-cost game. irobot-md irobot-trr
2004–2008	SIRIUScFraunhofer IFF · DE	Fully automated gantry + facade robot for an 80 m, 4,000 m² facade	research	Showed that the gantry (positioning, power, safety) is half the system; no scale-up evidence found. sirius
2013–2026	TOKA / Komodo crawlers + CantileverGecko Robotics · US	Wall-climbing inspection robots and an analytics platform for power, refining and naval assets	active	Data platform plus robots won defence contracts ($71M, expanded to a $300M IDIQ) and a $1.25B valuation; the money is in the data layer, not the robot alone. gecko-trr gecko-site
2013–2026	Relay (ex-Savioke)Relay Robotics · US	Hotel and hospital delivery robot	active	Savioke's assets were acquired by an undisclosed buyer and relaunched as Relay Robotics in May 2022; hospitality stayed the core market, with hospitals (10 at launch) and real estate as adjacent markets. Single-task delivery needed new verticals to grow. relay-trr
2014–2026	E4 / T4 / H4 waterless PV cleaningEcoppia · IL	Row-captive nightly microfibre + airflow PV cleaning robots	active	The most scaled fully autonomous maintenance robot found: captive to fixed geometry, so perception is minimal. Template for ORIJO surface robots. ecoppia-site ecoppia-ketura yt-ecoppia
2014	Ketura solar parkEcoppia / Arava Power · IL	100 robots clean a 20-acre PV park nightly without water	shipped	Replaced roughly nine water washes a year (each up to five days) with zero-water nightly cleaning. ecoppia-ketura
2014	Ecuador irrigation canalTU Delft / partners · EC	First field application of bacterial self-healing concrete with natural fibres	research	Field use in an Andean canal where about 70% of water was lost through unlined walls. jonkers-ecuador
2015–2026	Henn-na HotelH.I.S. Group · JP	Robot-staffed hotel with 243 robots	active (scaled back)	Decommissioned about half of 243 robots in 2019 because they created more work for humans; the chain survived by keeping robots as novelty/check-in features with human staff behind them (25 locations listed in Oct 2026). Guest-facing robots without a human backup fail; robots as an attraction plus humans as the service can persist. henn-futurism henn-verge henn-bhimasta henn-site
2015–2026	Basilisk self-healing concreteBasilisk (TU Delft spin-off) · NL	Bacteria-based healing agent, repair mortar and liquid repair	active	Commercial for sub-millimetre cracks in water-exposed concrete; field projects claimed. basilisk wiktor
2015–2016	HobbitTU Wien et al. · AT	Assistive manipulation robot in 16 private homes	research	Real homes exposed misinterpretation failures that never appeared in labs. hobbit
2015–2018	SYMPARTNERTU Ilmenau / MetraLabs · DE	Companion robots in 20 senior households, one week each, unsupervised	research	Seniors adapted quickly; robustness under domestic conditions was the hard part. sympartner
2016–2026	ANYmal / ANYmal XANYbotics (ETH Zurich spin-off) · CH	Autonomous quadruped inspector with thermal, acoustic and gas payloads; Ex-proof variant	active	Offshore HVDC field trials showed legged robots can patrol human-built plants; product runtime still ~90 min, so docks are essential. anymal-spec anymal-hvdc anybotics-60m
2017–2026	ServiBear Robotics (LG Electronics 51%) · US/KR	Restaurant and hospitality serving robots	active	Consolidated into a large manufacturer (LG call option to 51% exercised 2025–01–22); KED Global reports it sounding out PE firms for up to 400 billion won (~$300M) pre-IPO funding ahead of a potential Nasdaq listing (2026–09–07, reported, not confirmed by the company). bear-rave bear-ked
2017–2020	MoveCareEU H2020 consortium · EU	Assistive robots in 10 older adults' apartments for 130 weeks	research	Remote supervision increased acceptance. Long-term autonomy needs a human channel. luperto
2019–2026	OzmoSkyline Robotics · US/IL	KUKA arm on a building maintenance unit cleaning a 45-storey Manhattan facade	pilot	Claims ~3x speed but still needs a rooftop supervisor; lidar struggles on transparent glass, so ultrasonic sensors were added. ozmo-trr yt-ozmo
2019–2026	AstrobeeNASA Ames · US	Three free-flying IVR robots with dock recharge on ISS	active	>165 operations; ~80% of console time without crew; biggest problem was localization in a changing interior; too many ORUs hurt the design. astrobee-nasa astrobee-tfr
2019–2023	ISAACNASA · US	Integrated autonomous caretaking: robots + spacecraft systems + ground	research	Frames caretaking for uncrewed phases; Gateway planned to be uncrewed >85% of the time. isaac astrobee-tfr
2020–2026	SpotBoston Dynamics (Hyundai) · US	Commercial quadruped for inspection, data capture and manipulation	shipped	>1,500 units in customer hands (claimed); AB InBev cut average repair times from months to 13 days (claimed). Value comes from repeatable routes feeding work orders. spot-spec spot-retro
2020–2026	Stretch 3Hello Robot · US	Lightweight wheeled mobile manipulator	shipped	Sold to research ($24,950); vendor says consumer home market is not ready yet. Shows a low-mass, low-force design that is safe in homes. stretch-spectrum
2020–2029+	Canadarm3 / GatewayCSA / MDA / NASA · CA/US	8.5 m arm plus dexterous arm and ORU/tool caddy, some tasks autonomous	research	Delivery no earlier than 2029; CSA states intent to repurpose investments to the next lunar phase, so the Gateway path is uncertain. canadarm3 gateway-nasa
2023–2026	NEO1X Technologies · NO/US	Home humanoid sold at $20,000 or $499/month with remote 'Expert Mode' teleoperation	pilot	Teleoperation openly offered as the bridge to autonomy; WSJ found it mostly teleoperated. Honesty about remote operators is now a product feature. neo-newatlas neo-kalil yt-neo-wsj
2023–2026	Large Behavior ModelsToyota Research Institute · US	Multitask diffusion-transformer policies pretrained on ~1,700 h of demos	research	Finetuning needs under 30% of task data in sim to match single-task (interpolated estimate); the paper cites ~1,800 controlled real-world trials (elsewhere ~2,700 rollouts) to evaluate rigorously. Published in Science Robotics (2026–04–15). lbm
2023–2026	DigitAgility Robotics · US	Bipedal tote-handling humanoid at GXO, Schaeffler, Amazon, Toyota	active	>65,000 operating hours; Digit 5 claims a 9-minute charge. Fast charging changes how many robots a job needs. digit5
2025–2026	Helix / Figure 03Figure AI · US	Dual-system VLA (7B planner at 7–9 Hz, 80M controller at 200 Hz) for humanoids	pilot	~500 h of teleop data trained a generalist upper-body controller; no public success-rate numbers. figure-helix
2025	pi0.5Physical Intelligence · US	VLA co-trained on heterogeneous robots and web data, tested cleaning unseen homes	research	Open-world generalisation to new homes is now demonstrated in research; reliability for unattended use is not. pi05
2026	Walden RoboticsWalden Robotics (origin claim 'spun out of a Toyota lab' unverified) · US	General-purpose industrial robots; $300M seed	unverified	Headline only: a Business Wire release (2026–07–15) title says it launched with $300M; the page returned 403 and the round type, investors and founders could not be verified. Do not cite as evidence. walden

Failures, read for lessons

Henn-na Hotel decommissioned about half of its 243 robots (2019)
WhyGuest-facing robots handled only scripted cases (foreign passports, local questions); luggage robots reached about a quarter of rooms; failures created more human work.
Lesson for AnygroundKeep robots back-of-house and out of guest moments; budget a human runner for the hotel; never deploy a robot whose failure mode is a guest waiting.
iRobot filed Chapter 11 (Dec 2025) and was taken over by its contract manufacturer
WhyA narrow single-task consumer robot commoditised; the Amazon acquisition collapsed in 2024 and competitors undercut it.
Lesson for AnygroundDo not sell the robot; sell uptime of the building (lease/service). Robots are cost lines inside the SHIKINEN lease.
Savioke relaunched as Relay Robotics after acquisition (2022)
WhyUndisclosed acquirer formed a new entity; hotel delivery alone was a small market, so it expanded into hospitals and real estate while keeping hospitality as core.
Lesson for AnygroundSingle-task hotel robots need long corridors and volume; a 6-key ORIJO hotel has neither.
1X NEO revealed as heavily teleoperated in the WSJ test (2025)
WhyGeneral home autonomy is not ready; the company openly uses remote operators.
Lesson for AnygroundDesign for teleoperation from day one, with consent, privacy zones and logged interventions; market honest 'robot plus remote human' service.
Astrobee: too many components designed as ORUs
WhyCaptive fasteners, access geometry, mass and size grew; team judged that a few larger replaceable sub-assemblies would capture most of the benefit.
Lesson for AnygroundSwap at organ level (one JO-PORT organ = one ORU) and at wear-part level only.
Astrobee: modularity broken by unspecified interface properties (EMI on power bus, nozzle noise amplified by plenum)
WhyPower-quality and acoustic requirements were never defined at the module interface.
Lesson for AnygroundJO-PORT spec must include DC ripple/EMI limits, acoustic and thermal limits, not just geometry and connectors.
Astrobee: localization was the biggest continuing challenge
WhyThe ISS interior changes constantly, so maps went stale faster than planned.
Lesson for AnygroundUse fixtures (datum sockets, rails, docks) as the map; never rely on visual maps of rooms that guests rearrange.
ISS ECLSS: some ORU failure rates were badly underestimated (one ~7.8x)
WhyPre-flight reliability predictions lacked operating data.
Lesson for AnygroundStart each organ type with pessimistic priors, log every failure across the fleet, and size site spares from posteriors.
Hotel staff shelve robots after adoption (robot usage resistance)
WhyRobot-related extra workload, low usability, techno-insecurity.
Lesson for AnygroundMeasure staff minutes added by robots, not only saved; a robot that adds work is removed.
Gateway/Canadarm3 programme uncertainty (2025–2026)
WhyArtemis architecture changes; CSA states intent to repurpose Canadarm3 investments.
Lesson for AnygroundDo not anchor the TSUKI roadmap to a single agency programme; reuse terrestrial service grammar.
Ozmo still needs a rooftop supervisor; lidar failed on glass
WhyTransparent surfaces defeat lidar; safety at height requires a human.
Lesson for AnygroundAdd contact or ultrasonic sensing for glazing (YOSHI-6G); keep facade work on rails, not free-flying.

文04 Key papers

The science underneath, 29 papers.

Newest first. Each line states the finding that the dossier relies on.

2026
Astrobee: Free-Flying Robots for the International Space Station
Alexandrov, Barlow, Bualat, Fong, Smith et al. · IEEE Trans. Field Robotics · field-trial
Over 1,200 h of console support needed only ~220 h of crew (~80% crew-free); 4 hardware problems across ~379 robot and dock sessions.
2025
A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation
TRI LBM Team (Barreiros et al.) · arXiv 2507.05331 / Science Robotics 2026 · experiment
Finetuned LBMs match single-task performance on unseen sim tasks with under 30% of the task data (interpolated estimate), using about 1,695 h of pretraining demos (~545 h TRI + ~1,150 h OXE).
2025
pi0.5: a Vision-Language-Action Model with Open-World Generalization
Physical Intelligence (Black, Brown et al.) · arXiv · experiment
Co-training on heterogeneous robot and web data lets the model do multi-step tasks in homes not seen in training (no single headline number).
2024
DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset
Khazatsky, Pertsch, Nair et al. · RSS · experiment
The dataset holds 76k demonstrations collected in diverse real environments, improving policy robustness.
2023
Robots in Inspection and Monitoring of Buildings and Infrastructure: A Systematic Review
Halder, Afsari · Applied Sciences · review
A content analysis of 269 papers found nine robot types, with UAVs most common, then UGVs.
2023
PV module vibration by robotic cleaning
Figgis, Bermudez, Garcia · Solar Energy · field-trial
Field measurement at Qatar's QEERI of module vibration induced by cleaning robots (full text not accessed; values not verified).
2023
Diffusion Policy: Visuomotor Policy Learning via Action Diffusion
Chi, Xu, Feng et al. · RSS / IJRR · experiment
Across 15 tasks from 4 benchmarks (IJRR version; 12 tasks in the RSS 2023 version), it averages 46.9% better than the prior state of the art.
2023
Advances in Space Robots for On-Orbit Servicing: A Comprehensive Review
Ma, Jiang, Liu, Xie · Advanced Intelligent Systems · review
Names five key challenges, including multifunctional end-effectors and ground teleoperation with long delays.
2022
The perils of hotel technology: The robot usage resistance model
Fu, Zheng, Wong · Int. J. Hospitality Management · model
Staff resistance comes from excessive robot-related workload, techno-insecurity and low usability, so robots get shelved after adoption.
2022
User feedback and remote supervision for assisted living with mobile robots: long-term autonomy field study
Luperto, Romeo, Monroy et al. · Robotics and Autonomous Systems · field-trial
A 130-week study in 10 older adults' apartments found remote supervision increased acceptance.
2022
A Digital Twin predictive maintenance framework of air handling units based on AFDD
Hosamo, Svennevig, Svidt, Han · Energy and Buildings · field-trial
Combines APAR fault rules, ML condition prediction and maintenance planning, validated on Aug 2019-Oct 2021 data from a Norwegian building.
2022
Application of an adapted FMEA framework for robot-inclusivity of built environments
Ng, Yeo, Ng, Budig et al. · Scientific Reports · model
Most service-robot failures trace to poor furniture design, missing clearances and hazard indicators, and sub-optimal interior planning.
2022
A review on self-healing polymers for applications in spacecraft and construction of roads
Goyal, Agarwal, Bhatnagar · J. Applied Polymer Science · review
Compares the healing efficiency and limits of extrinsic and intrinsic self-healing polymers for space and roads.
2021
ANYmal in the Field: Solving Industrial Inspection of an Offshore HVDC Platform with a Quadrupedal Robot
Gehring, Fankhauser, Isler et al. · Field and Service Robotics (FSR 2019), Springer Proceedings in Advanced Robotics 2021 · field-trial
First North Sea HVDC platform trial of autonomous legged inspection (abstract only; numbers not verified).
2021
The Robot Household Marathon Experiment
Kazhoyan, Stelter, Kenfack et al. · ICRA · experiment
Tested extended fully autonomous pick-and-place and container opening in a household and reported the limits hit.
2021
ISAAC: An integrated system for autonomous and adaptive caretaking
Smith, Bualat et al. · ISS R&D Conference · experiment
Uses Astrobee and Robonaut as analogues to integrate IVR robots with spacecraft systems for uncrewed caretaking (no metrics given).
2020
Her Name Was Chloe and She Delivers Stuff to Your Room
Salminen, Milenkovic · NordiCHI · field-trial
Of 1,289 TripAdvisor reviews of six robot hotels, 92% were positive, mostly about novelty; failures show human backup is needed.
2020
Scaled Autonomy: Enabling Human Operators to Control Robot Fleets
Swamy, Reddy, Levine, Dragan · ICRA · experiment
Learning which robot most needs a human lets one operator supervise more robots than unaided attention allows.
2019
Techno-Economic Assessment of Soiling Losses and Mitigation Strategies for Solar Power Generation
Ilse, Micheli, Figgis et al. · Joule · model
Even with optimized cleaning, soiling cuts global solar output by at least 3–4%, worth 3–5 billion EUR a year in 2018, projected to rise to 4–7% by 2023.
2019
What causes the adoption failure of service robots? A case of Henn-na Hotel
Bhimasta, Kuo · UbiComp/ISWC Adjunct · field-trial
A thematic analysis of 250 negative reviews found six failure themes, including the need for human intervention, usefulness and embodiment.
2019
Results of Field Trials with a Mobile Service Robot for Older Adults in 16 Private Households
Bajones, Fischinger, Weiss et al. · ACM THRI · field-trial
All 16 users interacted with Hobbit daily over multiple weeks; real homes exposed misinterpretation failures.
2019
Living with a Mobile Companion Robot in your Own Apartment: 20-week field study with 20 seniors
Gross, Scheidig, Muller et al. · ICRA · field-trial
Two robots served 20 single-person households (ages 62–94) for a week each with no on-site support.
2019
On-orbit service (OOS) of spacecraft: A review of engineering developments
Li, Cheng, Liu et al. · Progress in Aerospace Sciences · review
Reviews OOS from Hubble's five servicing missions through ETS-VII and Orbital Express (384 citations).
2018
Review of UAS applications in the built environment: towards automated building inspection procedures using drones
Rakha, Gorodetsky · Automation in Construction · review
Proposes a standard drone IR energy-audit procedure, tested on the Syracuse University campus (465 citations).
2018
International Space Station Operational Experience and its Impacts on Future Mission Supportability
Owens, de Weck · ICES-2018–198 · model
18% of ECLSS ORUs fail more often than predicted (6% by more than 4x) and 26% are twice as reliable; corrected rates save 3.9–6.0 t of Mars spares.
2013
Climbing robots for maintenance and inspections of vertical structures: a survey
Schmidt, Berns · Robotics and Autonomous Systems · review
Landmark survey of adhesion and locomotion principles for climbing robots (282 citations); not read in full here.
2011
Quantification of crack-healing in novel bacteria-based self-healing concrete
Wiktor, Jonkers · Cement and Concrete Composites · experiment
Landmark bacterial healing study (1,177 citations); the healed crack-width values were not verified here.
2008
Kinematics, sensors and control of the fully automated facade-cleaning robot SIRIUSc
Elkmann, Lucke, Kruger et al. · Industrial Robot · field-trial
A fully automated gantry plus robot system was built for an 80 m, 4,000 m² facade.
2006
Crew maintenance lessons learned from ISS and considerations for future manned missions
Bertels · SpaceOps 2006 · field-trial
Search snippet says ISS crews had spent over 4,000 hours on maintenance by 2006 (full text not accessed; treat as unverified).

数05 Numbers ledger

Every number, with its basis.

Measured, reported, claimed, calculated or target: the basis travels with the value.

Quantity	Value	Unit	Basis	Sources
Global solar output lost to soiling even with optimized cleaningWorth 3–5 billion EUR a year; sets the value of PV cleaning robots	3–4 (4–7 projected for 2023)	% of production	calculated	ilse-joule
Ecoppia cumulative panels cleaned / sessionsCompany counters on 35+ sites and 5,000+ MW; an embedded figure says 4,000 MW (inconsistent)	3.89e9 / 10.51e6	panels / sessions	claimed	ecoppia-site
Ketura robot coverage rate (per robot)Times of Israel: each robot covers about 100 ft2 of panel per minute; 100 robots clean the 20-acre park nightly with zero water, replacing about nine water washes a year (each up to five days)	~100 (about 9.3)	ft2/min (m²/min)	reported	ecoppia-ketura
Dust efficiency loss quoted by Ecoppia (2014)Upper-bound desert claim; contrast with the 3–4% global average in NUM-01	up to 35	%	claimed	ecoppia-ketura
Spot runtime / recharge / payloadIP54, -20 to 55 C; run:charge about 1.5:1	90 / 60 / 14	min / min / kg	reported	spot-spec
Spot units in customer handsCompany page 2026	>1,500	robots	claimed	spot-spec
Spot industrial assets inspected in 2022AB InBev: average repair time from 'a few months' to 13 days	>130,000	assets	claimed	spot-retro
ANYmal battery runtime and quick charge10 kg extra payload; IP67; customer claims +1.5% production uptime	90 run / 100 to 70% / 180 full	min	reported	anymal-spec
Digit 5 charge time / runtime / run:chargeDigit 4 for comparison: 1–2 h charge, ~4 h runtime, 2:1. Digit 4: >65,000 operating hours at customer sites and >100,000 totes at ~98% task accuracy at GXO Flowery Branch	9 min; ~90 min; 10:1	min; min; ratio	claimed	digit5
Gecko Robotics valuation / Series DSeries D and valuation reported by The Robot Report 2025–06–12. Navy figures are company press releases (claimed): $71M deal 2026–03–17; IDIQ expanded to $300M 2026–10–20 with 'savings up to $8M per ship'	1.25 / 0.125	USD billion	reported	gecko-trr gecko-site
Henn-na Hotel robots decommissionedLuggage robots could reach only about a quarter of rooms; in-room assistant woke guests on snoring	about half of 243	robots	reported	henn-futurism
Robot-hotel reviews that were positiveMostly novelty-driven; failures stress human backup	92	% of 1,289 reviews	measured	chloe
Ozmo speed vs manual window cleaning45-storey tower, rooftop supervisor still present	~3x	ratio	claimed	ozmo-trr
1X NEO price / mass / battery / noiseRemote 'Expert Mode' teleoperation; WSJ: mostly teleoperated during visit	20,000 USD or 499 USD/month; 30 kg; ~4 h; 22 dB	mixed	claimed	neo-newatlas neo-kalil
Helix model splitRuns on two onboard embedded GPUs	7B @ 7–9 Hz planner; 80M @ 200 Hz controller; ~500 h teleop data; 35 DoF	params / Hz / h	reported	figure-helix
LBM data needed to match single-task performance on unseen sim tasksInterpolated from sim data-fraction sweeps (paper's own wording), not a direct real-world measurement. Pretraining ~1,695 h (~545 h TRI + ~1,150 h OXE); real SetBreakfastTable: LBM with 15% of the data beat single-task with 100% (single-task empirical success was 0)	<30	% of task-specific data	calculated	lbm
Diffusion Policy average improvement15 tasks, 4 benchmarks (IJRR 2024 version; the RSS 2023 arXiv version reports 12 tasks with the same 46.9%)	46.9	%	measured	dp
Astrobee console hours vs crew hoursNearly 80% of console time without crew; >165 real-time operations, >90 guest-science sessions	>1,200 / ~220	h	measured	astrobee-tfr
Astrobee hardware problems per sessionTable X: Dock 164 sessions/2 problems, Batteries 164/0, Bumble 135/0, Honey 59/1, Perching Arms 23/0, Queen 21/1. 379 counts Dock + 3 robots only; sessions overlap across rows, so treat as an indicative rate, not an MTBF	4 / 379 (about 1.1); 4 / 566 (about 0.7) if all Table X rows are counted	problems / sessions (%)	calculated	astrobee-tfr
Astronaut time per SPHERES session (pre-Astrobee)Astrobee's dock-recharge, fan propulsion design was driven by removing this burden	~4	h	reported	astrobee-tfr
Planned Gateway uncrewed share (as cited by Astrobee team)Stated in the Astrobee TFR paper (its ref. [29]); NASA's Gateway page (2026–07–23) gives no percentage and notes Artemis architecture updates of Feb 2026, so the programme is uncertain	>85	% of time	reported	astrobee-tfr
ISS ECLSS ORU failure-rate surprisesBayesian updates of on-orbit data	18% worse; 6% >4x worse; 26% ≥ 2x better; max 7.8x worse	% of ORUs	measured	owens-iss
Spares mass saved by ISS-informed failure rates (1,200-day Mars mission)Also uncovered underestimates worth an order of magnitude in risk	3.9–6.0	t	calculated	owens-iss
Canadarm2 length / joints / max payloadWalks end over end between grapple fixtures (analogue for TSURU)	17.6 / 7 / 116,000	m / joints / kg	reported	mss-wiki
Canadarm3 large arm length; deliveryIncludes ORU and tool caddy; some tasks autonomous	8.5 m; NET 2029	m; year	reported	canadarm3
Self-healing concrete crack widthBasilisk; activation needs water	up to 1	mm	claimed	basilisk
Long-term home robot field study lengthRemote supervision raised acceptance	130	weeks (10 apartments)	measured	luperto
iRobot Chapter 11 filing; Picea unsecured claimCombined debt and assets 480.3 M USD at 2025–06–28; taken over by its contract manufacturer	2025–12–14; 98.9	date; USD million	reported	irobot-md
Stretch 3 price / massResearch market; vendor says consumer homes are not its market yet	24,950 / 24.5	USD / kg	reported	stretch-spectrum
Remote-operator capacity for ORIJO homesFormula in sizing_reference.home_4p; assumes 5% intervention rate at 5 min each (target, unvalidated)	~200	homes per operator FTE	calculated	astrobee-tfr scaled-aut
Spot Explorer kit list priceRobot, two batteries, charger, tablet, cases, SDK; US commercial/industrial sale only at launch; current pricing is by quote	74500	USD (2020)	reported	spot-slashdot

段06 Autonomy ladder

From manual to Earth-independent, in human minutes.

Human minutes per week for this layer at each level (log bars). The unified site-level scale is on the Autonomy Stack page.

L0
All inspection, cleaning and repair done by residents or tradespeople on call.
Manuals, tools
~120–180 (estimate for an off-grid home with PV, water treatment and farm; unsourced, to be measured) min/wk
L1
Sensors and alarms: fixed IoT sensors and a dashboard tell people what to do.
IoT sensors, rule-based FDD
~90 (estimate) min/wk
L2
Captive single-task robots: PV/glazing row robots, robot vacuum, scheduled cartridge reminders.
Ecoppia-class row robots, docks
~60 (estimate) min/wk
L3
Rail-captive service robot does inspections and organ/cartridge swaps at standard ports; remote operator handles exceptions.
TSURU, JO-PORT, failure ledger, remote ops centre
~20 on site + ~6 remote (target, see sizing) min/wk
L4
Mobile manipulator handles open-ended indoor tasks with learned policies; teleop fallback below 1% of tasks.
LBM/VLA policies, Stretch/humanoid, fleet learning
~10 on site + ~3 remote (target) min/wk
L5
Self-maintaining: robots swap organs, receive drone resupply and service each other; humans only for depot repair and renewal.
Heavy-lift drone resupply, robot-serviceable robots, self-healing seals
~0 on site; depot time per swapped organ (speculative) min/wk

ORIJO roadmap for this layer

2027 L1 · v3 Phase 1 Bench
2029 L3 · v3 Phase 3 Catalogue (first IE homes, HATA-MURA farms)
2032 L4 · v3 Phase 4 Extreme (KITA winter-over)

結07 Interfaces

What crosses to the other layers.

Flows into and out of this layer from the synthesis, coloured by currency.

形Structure →Datum sockets every 910 mm, crease rails, robot-readable latch flagsabout 8 per 7.28 m ridge line sockets per YOSHI-4 ridge calculated
力Energy →Robot charging on JO-PORT docksabout 0.25 (TSURU-class patrol at 60 W) kWh/week calculated
水Water →Cartridge swaps (filters, UV, membranes, sorbent, pump tubing)about 1.5 (12 consumables, mean 8-week interval) swaps/week target
環Loop →Sealed waste cassettes to swap (zeolite, BSF trays, compost bins)12–25 kg zeolite/week calculated
農Grow →Rail arm and cassette service; shared dock and rail profile30–45 at L3 (target ≤ 30) human min/week per cassette target
建Build →Installs rails, sockets, docks; hands over as-built point cloudone survey per deployment - design rule
→ 形StructureLatch audits, seal and hinge thermal scans, leg re-levelling2 × 15 min patrols/day target
→ 力EnergyPV cleaning (recovered yield), battery/inverter organ swaps3–4 global soiling loss average (more in deserts) % of PV yield calculated
→ 水WaterFilter/membrane/UV/sorbent cartridge swaps10 robot min/swap target
→ 環LoopZeolite and BSF cassette swaps; odour patrol (NH3/VOC)1 (zeolite cartridge 20–25 kg) swaps/week target
→ 農GrowRail-arm and UV-bar upkeep, tray jam clearingincluded in 30–45 per cassette human min/week target
→ 建BuildRe-levelling, refold before storms, depot repair of swapped organs0.5 with organ swap vs 24–72 with on-site repair MTTR h target
→ 心MindIntervention logs as demonstrations; failure ledger; teleop sessionsabout 6 remote min/week per home calculated
心Mind →Work orders, spare forecasts, remote-assistance routingabout 200 at 6 remote min/week; 40–160 at L3 15–60 min homes per operator FTE calculated

折08 ORIJO design

The recommended stack and what we would build next.

What ORIJO adopts from the state of the art, the new mechanisms this layer adds and the 12-month prototype for each, with its pass and kill criteria.

Subsystem	ORIJO choice	Why
CARE-S1 Sense and patrol	Fixed sensors at every JO-PORT + TSURU-mounted thermal/acoustic head on crease rails; rent legged inspectors per cluster visit	Rail-captive gives repeatable viewpoints without mapping; legged robots are COMMERCIAL but too costly per home.
CARE-S2 Surface keeping	Ecoppia-style waterless helix-microfibre crawler on MIURA/BYOBU rows, docked at hinge ends	Most proven autonomous maintenance robot; zero water fits desert sites.
CARE-S3 Swap-in maintenance	Organ-level ORUs at JO-PORT + wear-part cartridges with one handle standard	ISS/Dextre shows ORUs enable robot maintenance; Astrobee shows the right granularity is coarse.
CARE-S4 Mobile manipulation	No humanoid in v1; wheeled Stretch-class manipulator only in KURA and farm bays from the Catalogue phase; hotel housekeeping stays human	Home autonomy is still teleop-heavy (NEO); Henn-na shows guest-facing failure costs.
CARE-S5 Prognostics	Fleet Bayesian failure ledger (DAICHO) + rule-based FDD for HRV/heat pump	Identical organs across sites make posteriors converge fast; ISS saved tonnes of spares this way.
CARE-S6 Remote ops	MADOGUCHI remote ops desk with goal-level commands and intervention logging	Astrobee: ~80% of operations need no person on site; logged interventions become training data.
CARE-S7 Passive durability	Bacterial self-healing concrete for ASHI pads and tanks; sacrificial wipers/seals elsewhere	Cheap reduction in task count; self-healing polymers are not yet proven outdoors.
CARE-S8 Docks	JO-PORT docks on DC spine; prefer fast-charge chemistries	Run:charge ratio sets robot count; Digit 5 shows 10:1 is possible.

01 New mechanism

KAWARI handle (替わり, 'replacement')

One robot-graspable handle and lead-in standard for every swappable organ, cartridge and robot wear part across water, energy, grow and loop layers.

How it works

A 60-degree cone pin plus a flat-faced T-handle sits on the cartridge's front face. TSURU or a person pushes it into a socket whose cone gives about 15 mm capture, after which two balls seat in V-grooves to fix position and a quarter-turn captive bolt preloads it. Dry-break couplers mate only after the datum is seated. An NFC tag in the handle carries the passport ID, and a reed switch confirms lock. The design follows ISS ORU/grapple practice at organ granularity, not component granularity.

12-month prototype: 3 socket variants (water filter, sorbent tray, battery module dummy) and 20 cartridges; a UR-class arm and a human test both
BOM: USD 8,000–12,000 (estimate)
Test: 2,000 insert/remove cycles by arm with ± 10 mm and ± 3 deg initial error; 200 human swaps by untrained users; leak test at couplers
Pass: ≥ 99.5% first-try seat by arm; median human swap ≤ 90 s; zero coupler leaks at 3 bar
Kill: <98% arm seat rate after two design iterations, or human swap >3 min

02 New mechanism

HOKI crease broom (箒, 'broom')

Waterless PV and glazing crawler that rides the MIURA wing hinge rail, parks in a JO-PORT dock at the wing root.

How it works

A helical microfibre cylinder (Ecoppia/Evermore principle) rotates while the carriage travels along the fold line, generating a directional airflow that pushes dust to the low edge. The rail is the hinge line, so the crawler needs only end-stops and a gap sensor. Because MIURA wings fold, the crawler docks before folding and the dock doubles as its storm shelter. Cleaning is triggered when a reference cell shows a soiling ratio below a threshold.

12-month prototype: One 6 m MIURA-wing mock-up with 6 PV modules and reference cell; crawler with 1.8 m helical microfibre cylinder
BOM: USD 6,000–9,000 (estimate)
Test: Desert dust deposition chamber or field site, 90 nights; measure soiling ratio, module vibration (accelerometer) and brush wear
Pass: Keeps soiling loss <2% between cycles; module acceleration below manufacturer limit; brush life ≥ 6 months
Kill: Soiling loss >4% or visible microcracks in EL imaging after 90 nights

03 New mechanism

DAICHO ledger (台帳, 'register')

A fleet-wide Bayesian failure-rate ledger and spares planner for every ORIJO organ passport.

How it works

Each organ type starts with a pessimistic prior failure rate and error factor (ISS MADS practice). Operating hours without failure and each failure update a gamma-Poisson posterior. Site spares are sized so that the probability of sufficiency over the resupply interval is at least 99%, and swap intervals move with the posterior. Remote-operator interventions and robot task failures are logged in the same ledger, so robots get MTBF too.

12-month prototype: Software on existing telemetry from 2 pavilion sites plus bench rigs; import ISS-style priors
BOM: USD 15,000–30,000 (engineering time, estimate)
Test: Back-test on bench failure data; compare spares predicted vs consumed over 12 months
Pass: Predicted spares cover ≥ 99% of demand with <30% excess mass
Kill: Insufficient failure events to update priors (keep as a manual log until the fleet is >20 sites)

04 New mechanism

MADOGUCHI desk (窓口, 'service window')

Remote operations centre where one operator supervises many sites' robots with goal-level commands.

How it works

Robots stop in a safe state below a confidence threshold and post a ticket with video and a suggested action. A scaled-autonomy scheduler ranks tickets by urgency and expected operator time. The operator issues goal-level commands (pose, which cartridge, abort) rather than joystick control, which tolerates satellite latency. Every resolved ticket is stored as a demonstration for policy finetuning, and residents can see and veto camera access through privacy zones.

12-month prototype: One desk, 3 test robots (TSURU mock, HOKI, rented quadruped), Starlink link
BOM: USD 20,000–40,000 excluding robots (estimate)
Test: 8 weeks of 24/7 operation across 2 sites; log interventions, latency, resolution time
Pass: Intervention rate ≤ 5% of tasks; median resolution ≤ 5 min; zero privacy-zone violations
Kill: Interventions >20% of tasks after 8 weeks, or operator load >30 min per site per week

05 New mechanism

MIMAMORI head (見守り, 'watching over')

A 3 kg sensor head (thermal camera, ultrasonic microphone, VOC/NH3 gas, RGB) carried by TSURU along ridge rails twice a day.

How it works

At each datum socket the head captures a fixed-pose thermal and acoustic frame. Because the pose repeats to within socket tolerance, frame-to-frame differencing reveals new hot joints, leaks and latch gaps without a learned map. Anomalies above baseline create a DAICHO ticket.

12-month prototype: Head on a manual rail trolley over a YOSHI mock-up with seeded faults (heated connector, 0.5 L/h leak, loose latch)
BOM: USD 4,000–7,000 (estimate)
Test: 60 days, 2 patrols/day, randomised seeded faults
Pass: Detects ≥ 90% of seeded faults within 24 h with <1 false alarm per week
Kill: <70% detection or >5 false alarms per week

Sizing reference

home 4p

formula: H_site = sum(tasks_human x min) ; R_robot = patrols/day x 7 x t_patrol + swaps/week x t_swap + PV passes/week x A_pv / r_clean ; H_remote = N_tasks x p_int x t_int ; homes_per_operator = (2,400 min/week x utilisation) / H_remote
basis: calculated from targets; only r_clean is a reported external number

hotel 6key

formula: turnovers/week = keys x occupancy x 7 / mean_stay ; H_housekeeping = turnovers x t_turn ; plus home-style care stack for the building
basis: calculated from assumptions; housekeeping time is the dominant term and must be measured at the Pavilion phase

farm module

formula: H = HATA robot service + cassette swaps + inspection exceptions ; robot uptime A = MTBF / (MTBF + MTTR)
basis: calculated; the point is that swap-based MTTR, not MTBF, dominates availability at remote sites

問09 Open problems

What is not solved yet.

No public long-run reliability data (MTBF, interventions per hour) for home mobile manipulators or humanoids; vendors publish demos, not failure logs.
Rail-captive robots in icing, snow load and sand: no field data for crease-rail locomotion (TSURU) in E3/E4/E5.
PV cleaning-induced module vibration and microcracks: the trade-off between cleaning frequency and module life is not quantified for lightweight folding PV wings.
Hotel room turnover remains human; bed-making and bathroom cleaning robots are not commercial. Real minutes per turnover must be measured in the Pavilion phase.
Teleoperation economics and privacy: the cost per remote intervention, latency tolerance over satellite links, and consent models for cameras in homes and guest rooms.
Interface specifications beyond geometry (EMI, ripple, acoustic, thermal) for JO-PORT, the gap Astrobee hit.
Self-healing polymers for outdoor hinge membranes: UV, cold-cycling and multi-cycle healing are unproven.
Fleet failure ledger needs enough identical units (>20 sites) before posteriors beat priors; early sites will run on pessimistic guesses.
Regulatory status of unattended drone inspection and heavy-lift resupply varies by country.
Gateway's uncertainty removes the main near-term public testbed for IVR caretaking; terrestrial analogues (ORIJO polar stations) may be the better testbed.

検10 Verification

Checked adversarially.

Claims re-checked

Confirmed

Corrected

Unverifiable

Substantially sound. No fabricated sources or DOIs found; numbers match primary sources with correct units. Corrections were about status currency (Henn-na still operating, iRobot restructured not dead), one overstated pivot (Relay), one unsupported detail (Ozmo sill sensors), one basis downgrade (LBM <30% is an interpolated sim estimate), one maturity downgrade (Astrobee ground ops PILOT 8), and the Walden item, which stays unverified.

Corrections made (19)

NUM-03~~Ketura robot coverage rate (ambiguous whether park or robot rate)~~ → per-robot rate, ~100 ft2/min (~9.3 m²/min); sizing r_clean is therefore per robot
NUM-09~~9 min; 10:1~~ → 9 min charge; ~90 min runtime; 10:1 (Digit 4: ~4 h run, 2:1) — runtime added so the ratio is not read as longer endurance
NUM-10 context~~June 2025; US Navy IDIQ expanded to $300M with claimed savings up to $8M per ship~~ → Series D and valuation reported by The Robot Report 2025–06–12. Navy figures are company press releases (claimed): $71M deal 2026–03–17; IDIQ expanded to $300M 2026–10–20 with 'savings up to $8M per ship'
NUM-16 basis~~measured~~ → calculated (interpolated, sim only)
NUM-17 context~~15 tasks, 4 benchmarks~~ → 15 tasks, 4 benchmarks (IJRR 2024 version; the RSS 2023 arXiv version reports 12 tasks with the same 46.9%)
NUM-19~~From Table X: Dock 164 sessions/2, Bumble 135/0, Honey 59/1, Queen 21/1 (omitted Batteries and Perching Arms rows)~~ → 4 / 379 (about 1.1); 4 / 566 (about 0.7) if all Table X rows are counted
NUM-21 sources~~['src-astrobee-tfr', 'src-gateway-nasa']~~ → ['src-astrobee-tfr'] only — NASA Gateway page does not state >85%
PRJ-06 sensors~~Claims ~3x speed but still needs a rooftop supervisor; lidar fails on glass and needed extra ultrasonic and sill sensors.~~ → Claims ~3x speed but still needs a rooftop supervisor; lidar struggles on transparent glass, so ultrasonic sensors were added.
PRJ-08 status~~failed / 2015–2019~~ → active (scaled back) / 2015–2026: Henn na Hotel chain lists 25 locations with front-desk and dinosaur robots in Oct 2026
PRJ-09~~Savioke was acquired and relaunched as Relay in 2022 and pivoted toward hospitals; single-task hotel delivery was not enough on its own.~~ → Savioke's assets were acquired by an undisclosed buyer and relaunched as Relay Robotics in May 2022; hospitality stayed the core market, with hospitals (10 at launch) and real estate as adjacent markets. Single-task delivery needed new verticals to grow.
failures: Savioke why~~Hotel delivery alone did not sustain the business; pivot to hospitals.~~ → Undisclosed acquirer formed a new entity; hotel delivery alone was a small market, so it expanded into hospitals and real estate while keeping hospitality as core.
PRJ-10~~Consolidated into a large manufacturer (LG 51% in Jan 2025); pursuing ~$300M pre-IPO round (2026, not verified).~~ → Consolidated into a large manufacturer (LG call option to 51% exercised 2025–01–22); KED Global reports it sounding out PE firms for up to 400 billion won (~$300M) pre-IPO funding ahead of a potential Nasdaq listing (2026–09–07, reported, not confirmed by the company).
PRJ-11 status~~failed / 2002–2025~~ → restructured: emerged from Chapter 11 as Picea subsidiary (TRR 2026–01–23); Roomba line continues
PRJ-15~~Finetuning needs under 30% of task data in sim to match single-task; ~1,800 real-world trials needed to evaluate rigorously.~~ → Finetuning needs under 30% of task data in sim to match single-task (interpolated estimate); the paper cites ~1,800 controlled real-world trials (elsewhere ~2,700 rollouts) to evaluate rigorously. Published in Science Robotics (2026–04–15).
PRJ-28~~Spun out of a Toyota lab / active~~ → status unverified; Toyota origin and 'seed' removed as unverified
PAP-04 year~~2019~~ → 2021 (published; conference 2019)
CARE-S6 Ground-operator model maturity~~COMMERCIAL 9~~ → PILOT 8
CARE-S3 Canadarm2+Dextre trl_why~~Operational on ISS since 2001 (arm) and 2008 (Dextre).~~ → added caveat: operational agency system, not commercially sold
CARE-S5 Bayesian ledger trl_why~~Used operationally by ISS logistics and maintenance.~~ → Used operationally by ISS logistics (MADS); the method is mature but has no off-the-shelf product for building fleets, so ORIJO's DAICHO implementation is TRL ~4.

Could not verify (4)

item: PRJ-28 Walden Robotics $300M seed / Toyota origin · reason: Business Wire page blocked (403); layer SERP budget exhausted; only the URL title supports $300M
item: PAP-06 Figgis 2023 vibration values · reason: full text not accessed (already flagged)
item: PAP-22 Bertels >4,000 crew maintenance hours · reason: DOI resolves, full text not accessed (already flagged)
item: Brush-crawler PV robot specs (SolarCleano) · reason: no source in file

Certification route

C-C1international · ISO 13482 predates learned behaviours
ISO 13482:2014 (personal care robots), ISO 10218–1/-2:2025 (industrial robots)
Hazard analysis, speed/force limits, protective stops
For ORIJO: Rail-captive TSURU and tray shuttles are easier to certify than free-roaming humanoids: fixed paths, guarded zones.

典11 Sources

70 sources, linked.

Raw data: /d/layers/care.json. Search log and method: Evidence.

src-gecko-trrGecko Robotics reaches unicorn status with latest funding (The Robot Report, 2025–06–12) · news · read
src-gecko-siteGecko Robotics home page (news items: US Navy $71M deal, IDIQ expansion to $300M) · company · read
src-gecko-cnbcGecko Robotics raises $125 million in deal valuing startup over $1 billion (CNBC) · news
src-pat-geckoEP3974823B1 Inspection robot (Gecko Robotics, priority 2016) · patent
src-anymal-specANYbotics ANYmal product page · company · read
src-anybotics-60mANYbotics raises $60M to expand inspection robot deployments worldwide (The Robot Report, 2024–12–12) · news
src-anymal-hvdcGehring et al., ANYmal in the Field: Solving Industrial Inspection of an Offshore HVDC Platform with a Quadrupedal Robot (FSR 2019) · paper
src-spot-specBoston Dynamics Spot product page · company · read
src-spot-retroA Retrospective on Uses of Boston Dynamics' Spot Robot (Boston Dynamics blog, 2024) · company · read
src-ecoppia-siteEcoppia home page (fleet counters) · company · read
src-ecoppia-keturaWorld's first self-cleaning solar park in the Arava Valley (Times of Israel, 2014–03–25) · news · read
src-pat-evermoreWO2023181015A1 Waterless solar panel cleaning robot with helix shaped microfiber elements (Evermore United SA) · patent · read
src-pat-ecoppia-pidWO2019186347A1 Solar panel waterless cleaning system with PID recovery · patent
src-ilse-jouleIlse et al., Techno-Economic Assessment of Soiling Losses and Mitigation Strategies for Solar Power Generation, Joule 2019 · paper · read
src-figgis-vibFiggis, Bermudez, Garcia, PV module vibration by robotic cleaning, Solar Energy 2023 · paper
src-ozmo-trrSkyline Robotics deploys Ozmo window cleaning robot in New York City (The Robot Report, 2024–08–28) · news · read
src-siriusElkmann et al., Kinematics, sensors and control of the fully automated facade-cleaning robot SIRIUSc for the Fraunhofer headquarters, Industrial Robot 2008 · paper · read
src-henn-futurismJapan's Robot Hotel Just Laid Off Half Its Robots (Futurism, 2019–01–15) · news · read
src-henn-vergeJapan's robot hotel lays off half the robots after they created more work for humans (The Verge, 2019–01–15) · news
src-henn-bhimastaBhimasta & Kuo, What causes the adoption failure of service robots? A case of Henn-na Hotel in Japan (UbiComp Adjunct 2019) · paper · read
src-chloeSalminen & Milenkovic, 'Her Name Was Chloe and She Delivers Stuff to Your Room' (NordiCHI 2020) · paper · read
src-rurmFu, Zheng, Wong, The perils of hotel technology: The robot usage resistance model (IJHM 2022) · paper · read
src-relay-trrSavioke is now Relay Robotics (The Robot Report, 2022–05–02) · news · read
src-bear-raveLG Buys Most of Bear Robotics (rAVe Pubs, 2025–01–30) · news · read
src-bear-kedLG Electronics-owned Bear Robotics eyes up to $300 mn pre-IPO funding (KED Global, 2026–09–07) · news
src-irobot-mdRoomba maker files for Chapter 11 bankruptcy (Manufacturing Dive, 2025–12) · news · read
src-irobot-trriRobot emerges from Chapter 11 as restructured Picea U.S. subsidiary (The Robot Report, 2026–01–23) · news
src-neo-newatlas1X's first robot housekeeper is available to pre-order in the US (New Atlas, 2025–10–29) · news · read
src-neo-kalilIs 1X's $20K Humanoid Robot for Home Too Good to Be True? (mikekalil.com, 2025–10–29) · news · read
src-yt-neo-wsjI Tried the First Humanoid Home Robot. It Got Weird. (WSJ, YouTube) · video
src-figure-helixHelix: A Vision-Language-Action Model for Generalist Humanoid Control (Figure, 2025–02–20) · company · read
src-stretch-spectrumStretch 3 Brings Us Closer to Realistic Home Robots (IEEE Spectrum, 2024–02–15) · news · read
src-digit5Agility Robotics Unveils Digit 5 (AI Insider, 2026–09–15) · news · read
src-lbmTRI LBM Team, A Careful Examination of Large Behavior Models for Multitask Dexterous Manipulation (arXiv 2507.05331; Science Robotics 2026) · paper · read
src-dpChi et al., Diffusion Policy: Visuomotor Policy Learning via Action Diffusion (RSS 2023 / IJRR 2024) · paper · read
src-pi05Physical Intelligence, pi0.5: a Vision-Language-Action Model with Open-World Generalization (arXiv 2504.16054) · paper · read
src-droidKhazatsky et al., DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset (RSS 2024) · paper · read
src-scaled-autSwamy, Reddy, Levine, Dragan, Scaled Autonomy: Enabling Human Operators to Control Robot Fleets (ICRA 2020) · paper · read
src-lupertoLuperto et al., User feedback and remote supervision for assisted living with mobile robots: a field study in long-term autonomy (RAS 2022) · paper · read
src-hobbitBajones et al., Results of Field Trials with a Mobile Service Robot for Older Adults in 16 Private Households (ACM THRI 2019) · paper · read
src-sympartnerGross et al., Living with a Mobile Companion Robot in your Own Apartment: 20-week field study with 20 seniors (ICRA 2019) · paper · read
src-marathonKazhoyan et al., The Robot Household Marathon Experiment (ICRA 2021) · paper · read
src-astrobee-nasaNASA Ames Astrobee page · agency · read
src-astrobee-tfrAlexandrov et al., Astrobee: Free-Flying Robots for the International Space Station (IEEE Trans. Field Robotics 2026, NTRS 20260001396) · paper · read
src-isaacSmith et al., ISAAC: An integrated system for autonomous and adaptive caretaking (ISSRDC 2021, NTRS 20210020134) · agency · read
src-gateway-nasaNASA Gateway Space Station reference page (updated 2026–07–23) [no uncrewed-share figure on page] · agency · read
src-canadarm3About Canadarm3 (Canadian Space Agency) · agency · read
src-mss-wikiMobile Servicing System (Wikipedia; secondary summary of NASA/CSA data) · agency · read
src-yt-dextreDextre changes a pump on the International Space Station (CSA animation, YouTube 2015) · video
src-owens-issOwens & de Weck, International Space Station Operational Experience and its Impacts on Future Mission Supportability (ICES-2018–198, NTRS 20200002558) · paper · read
src-bertelsBertels, Crew maintenance lessons learned from ISS and considerations for future manned missions (SpaceOps 2006) · paper
src-oos-liLi et al., On-orbit service (OOS) of spacecraft: A review of engineering developments (Prog. Aerospace Sci. 2019) · paper · read
src-oos-maMa et al., Advances in Space Robots for On-Orbit Servicing: A Comprehensive Review (Adv. Intell. Syst. 2023) · paper · read
src-halderHalder & Afsari, Robots in Inspection and Monitoring of Buildings and Infrastructure: A Systematic Review (Appl. Sci. 2023) · paper · read
src-rakhaRakha & Gorodetsky, Review of UAS applications in the built environment: towards automated building inspection (Autom. Constr. 2018) · paper · read
src-hosamoHosamo et al., A Digital Twin predictive maintenance framework of air handling units based on AFDD (Energy & Buildings 2022) · paper · read
src-rifmeaNg et al., Application of an adapted FMEA framework for robot-inclusivity of built environments (Sci. Rep. 2022) · paper · read
src-schmidt-bernsSchmidt & Berns, Climbing robots for maintenance and inspections of vertical structures: a survey (RAS 2013) · paper
src-basiliskBasilisk self-healing concrete product site · company · read
src-wiktorWiktor & Jonkers, Quantification of crack-healing in novel bacteria-based self-healing concrete (Cem. Concr. Compos. 2011) · paper
src-jonkers-ecuadorSierra Beltran, Jonkers et al., Field application of self-healing concrete with natural fibres as linings for irrigation canals in Ecuador (2015) · paper · read
src-shp-reviewGoyal et al., A review on self-healing polymers for applications in spacecraft and construction of roads (J. Appl. Polym. Sci. 2022) · paper · read
src-yt-ecoppiaEcoppia H4: autonomous robotic cleaning for fixed tilt and single axis trackers (YouTube) · video
src-yt-ozmoMeet Ozmo: The World's First Robotic-armed Window-Cleaning Robot (YouTube) · video
src-yt-anymal-dockPrototype Docking Station for ANYmal (YouTube) · video
src-spot-dhlHyundai to buy 'tens of thousands' of Boston Dynamics robots (The Robot Report, 2025–04–04) · news
src-waldenWalden Robotics Launches with $300 Million (Business Wire, 2026–07–15) [page blocked (403); unverified] · news
src-v3-conceptAnyground ORIJO v3 concept.json (M04 JO-PORT, M09 TSURU, TSURU-1, KURA-3) · company · read
src-henn-siteHenn na Hotel official site (hotel list, technology section) · company · read
src-spot-slashdotSlashdot (16 Jun 2020) Boston Dynamics starts selling its Spot robot for $74,500 (Spot Explorer kit contents) · news · read

Care. Make the building the robot's map and fixture, swap whole organs, and keep a remote human for exceptions.

How the care layer works, option by option.

Sense and patrol (inspection)

Legged inspector (ANYmal, Spot)

Rail-captive sensor head (TSURU / crease rail)

Magnetic / climbing crawler (Gecko TOKA class)

Inspection drone

Intra-vehicular free flyer (Astrobee)

Surface keeping (PV, glazing, facade)

Row-captive waterless PV robot (Ecoppia class)

BMU-mounted robotic arm (Skyline Ozmo)

Brush crawler / hand-portable PV robot

Fixed-gantry facade robot (Fraunhofer SIRIUSc)

Swap-in maintenance (organ / ORU replacement)

Manipulator with tool changer on grapple fixtures (Canadarm2 + Dextre)

Organ-level JO-PORT swap (v3 M04)

Component-level ORUs everywhere

Mobile manipulation and housekeeping

Delivery cart robot (Relay, Bear Servi)

Wheeled single-arm manipulator (Hello Robot Stretch 3)

Home humanoid with teleoperation fallback (1X NEO)

Industrial humanoid (Agility Digit)

Learned policies (Diffusion Policy, TRI LBM, Helix, pi0.5)

Prognostics: digital twin and failure ledger

BIM/IoT digital twin with AFDD (Hosamo 2022)

Bayesian fleet failure-rate ledger (ISS MADS approach)

Acoustic and thermal anomaly mapping

Remote operations and teleoperation fallback

Ground-operator model (Astrobee/ISS)

Consumer teleop fallback (1X Expert Mode)

Remote supervision in assisted living (MoveCare field study)

Passive durability and self-healing materials

Bacteria-based self-healing concrete (Basilisk)

Self-healing polymers and coatings

Sacrificial wear parts by design

Robot docks and energy for the robots

Autonomous charging dock (Spot, ANYmal, Astrobee)

JO-PORT dock on DC spine

Who has tried this, and what happened.

Failures, read for lessons

Henn-na Hotel decommissioned about half of its 243 robots (2019)

iRobot filed Chapter 11 (Dec 2025) and was taken over by its contract manufacturer

Savioke relaunched as Relay Robotics after acquisition (2022)

1X NEO revealed as heavily teleoperated in the WSJ test (2025)

Astrobee: too many components designed as ORUs

Astrobee: modularity broken by unspecified interface properties (EMI on power bus, nozzle noise amplified by plenum)

Astrobee: localization was the biggest continuing challenge

ISS ECLSS: some ORU failure rates were badly underestimated (one ~7.8x)

Hotel staff shelve robots after adoption (robot usage resistance)

Gateway/Canadarm3 programme uncertainty (2025–2026)

Ozmo still needs a rooftop supervisor; lidar failed on glass

The science underneath, 29 papers.

Every number, with its basis.

From manual to Earth-independent, in human minutes.

What crosses to the other layers.

The recommended stack and what we would build next.

KAWARI handle (替わり, 'replacement')

HOKI crease broom (箒, 'broom')

DAICHO ledger (台帳, 'register')

MADOGUCHI desk (窓口, 'service window')

MIMAMORI head (見守り, 'watching over')

Sizing reference

home 4p

hotel 6key

farm module

What is not solved yet.

Checked adversarially.

Certification route

ISO 13482:2014 (personal care robots), ISO 10218–1/-2:2025 (industrial robots)

70 sources, linked.