Close and verify latches
After unfold, close each ridge splice latch, then confirm it with the Hall sensor. During commissioning, find leaks with the thermal camera while the KURA blower door pressurises the shell.
Latch driverM09 · Systems · Service robotEXPERIMENTAL · TRL 3
TSURU walks on the creases.
A 25–35 kg inchworm robot that locks its feet into datum sockets bolted to the ridge beam every 910 mm. The building is its map and its fixture, so a small robot can do precise work.
The building is the map. The socket is the fixture.
Service robots fail in buildings because buildings are unpredictable. ORIJO is predictable by construction. TSURU locks one foot, flips the other over to the next socket, locks again, and then works from a locked foot. The planner on the left runs a shift of versioned task contracts; every number on the stage is computed from the pose you see.
Fig. 01 · M09 TSURURidge rail · 8 sockets · 7.28 m
Phase—
Moment on locked foot0 N·m
Drag to orbit · click a part · playback about ×5 real time
Moment on the locked foot calculated
0N·m static
From the drawn pose: g · Σ m · lever, for a 30 kg body. Design peak ×1.7 = 0 N·m, against joints sized ≥ 400 N·m and locks rated ≥ 600 N·m target.
Capture error at the lead-in modelled
0.0mm
Each step arrives 14 mm off-centre on purpose. The 20–30 mm tapered lead-in pulls it home before the petals lock.
Travelled this shift
0.00m
0.91 m per step · 15–30 s per step = 0.03–0.06 m/s calculated.
Parts · click to inspect in 3D18
Unlock, arc, place, lock. Fifteen to thirty seconds, every 910 mm.
The free foot flips over the locked one, as NASA's ARMADAS inchworm robots do. Only one foot is ever free, and a person never has to trust the robot's sense of where it is. The socket tells it.
-
01
Unlock
The petals fold in, the lock sensor reads open, and the nose lifts clear of the cup. The other foot carries the whole robot through its lock: rated ≥ 0.6 kN·m and ≥ 3 kN fall-arrest target.
-
02
Arc
Two 550 mm links swing the free foot over the top to the next socket. It passes straight up at mid-swing, which keeps the lever on the locked foot short. With up to 8 kg at full swing, the peak is about 350–390 N·m calculated.
-
03
Place
The nose drops into a tapered cup, which pulls it onto the datum. ARMADAS petal feet capture 3.0 cm of in-plane misalignment measured, so TSURU's 20–30 mm lead-in means it never has to know exactly where it is.
-
04
Lock
The petals open under the cup's undercut and a sensor confirms the lock. Only then can the other foot let go. Sockets are bolted through to the ridge beam, not to the skin, and are drained so they clean themselves.
Rails
The rail sits on one panel and is broken at every transverse crease, so the vault can still fold. Rail continuity is a named risk.
Sockets
Stainless cups at 910 mm pitch, one per jo, are the same module the rest of ORIJO is laid out on.
Fall arrest
A continuous trolley line runs along every rail TSURU uses outdoors. If the robot falls, the line catches it, not a foot.
Coverage
One robot per contiguous rail network. A person carries it between modules in a 0.6 × 0.4 × 0.4 m case.
Why the spec changed: the statics did not close.
The first spec was a 25 kg body carrying 15 kg at full swing. That puts about 244 N·m on the locked foot before any dynamic factor. The adversarial review moved heavy payloads to double stance, with both feet locked. Try it.
Dynamic and slope factor
M = g · (mbody · L/2 + mpayload · L) · factor, with L = 0.91 m. These are the same formula and inputs as the concept dossier calculated.
Full swing allowed
387N·m peak at the locked foot
StaticPeak with factorJoint sizing ≥ 400Lock rating ≥ 600
A robot that measures itself drifts. One that measures against the structure doesn't.
NASA's ARMADAS assemblers walk on the lattice they build and position themselves relative to it. Their error does not pile up step after step. It settles toward the tolerance of the parts. We borrow the architecture, not the scale.
A · Dead reckoningFree-roaming robot, global coordinates. Every step adds its own scatter plus a small uncorrected bias.
B · Lattice-relative (TSURU)Each step targets the next socket. The lead-in captures it, and the error resets to the part tolerance.
A · error at last step
—mm
B · error at last step
—mm
0 re-tries (step missed the lead-in)
Precedent
SOLL-E is a 5-DoF bipedal inchworm. It made about 15,000 indoor moves on a 304.8 mm voxel lattice measured. Its successor weighs about 5.6 kg reported. TSURU is about five times heavier, because it works outdoors in wind and rain and carries cartridges.
What converges
In ARMADAS simulations, lattice RMS error falls toward the part error as the plates get thicker modelled. The toy model above shows the same idea. It is illustrative, not a TSURU measurement.
What does not
Accuracy is local. Socket-to-socket error is set by the socket and its lead-in. Hinge play adds several mm per joint, so building coordinates come from a survey after the vault is up, not from the robot.
Four jobs, all along the ridge. Everything else goes to someone else.
Tasks arrive as versioned contracts with a pass test. TSURU only takes work it can reach from the ridge rail. Narrow scope is the design: scope creep toward a general-purpose robot is listed as a risk, not a roadmap.
Scan hinge membranes
Sweep the crease membranes near the ridge with a thermal camera and log every frame to the module passport. A sensor trigger or the shikinen schedule turns a finding into a swap job.
Thermal cameraSwap filter, UV and sorbent cartridges
Swap cartridges through JO-PORTs and reload seed cartridges. The free foot can carry 5–8 kg at full swing. Up to 15 kg needs both feet locked.
GripperWipe the ridge strip
Clean sensor windows and facets inside one bay of the rail. A ridge robot cannot reach the eaves, so PV and slope washing go to a separate brush crawler or to people.
Wiper padWho does what on a site
| Work | Done by | Basis |
|---|---|---|
| Ridge-adjacent contracts (latch, scan, swap, wipe) | TSURU, ≥ 95 % autonomous completion | target |
| Exceptions inside a contract | Remote operator over Starlink | target |
| Anything outside a contract | Local human runner, plus a quarterly runner visit | target |
| PV and slope washing | Separate brush crawler, or people | target |
| Ground and cluster inspection | Rented legged inspector or farm utility rover, on scheduled visits | target |
| Cartridge resupply at roadless sites | Heavy-lift drone: 65 kg over 12 km on dual battery | claimed |
| Moving the robot between modules | A person, carrying a 0.6 × 0.4 × 0.4 m case | target |
>50%
of the Henn-na Hotel's 243 robots were retired measured. They had been put on stage, at the front desk, with the luggage and in the rooms. ORIJO takes the opposite brief.
Source: Engadget, 16 January 2019.
Robots on stage fail. Robots in the walls serve.Kage no omotenashi · invisible care
Never automated in ORIJO
- Greeting, check-in and anything a guest talks to
- Luggage, housekeeping theatre, in-room voice assistants
- Slopes, eaves and PV washing beyond the ridge's reach
- Any job without a versioned contract and a pass test
Automated, out of sight
- Latch close and verify along the ridge
- Thermal scans of hinge membranes
- Filter, UV, sorbent and seed cartridge swaps
- Logs to the module passport, with local-first firmware
The rule: autonomy lives back-of-house. Guests see doors, light, water and food, not robots. Every transformation also works by hand, so TSURU is an add-on and never the only way to do a job.
TSURU v0 on a 7.28 m bench rail, then on the half-shell.
- Build
- TSURU v0 on a 7.28 m bench rail and the 3.64 m rail of the M02 half-length test shell, with petal-lock feet, bolted sockets and a continuous fall-arrest trolley. Cable kit moves to month 12–24.
- Budget
- USD 40,000–80,000 BOM plus about 300,000–500,000 labour (2–3 FTE)
- Test
- Human baseline times measured first; then 200 trials each of latch verify and filter-cartridge swap; wind, rain, grit and ice in sockets; fall-arrest drop tests.
- Pass
- ≥ 95% autonomous completion; zero dropped payloads; zero failed locks under load; mean task time within 2x the measured human baseline.
- Kill
- < 80% completion after redesign: TSURU becomes a human-carried tool and rails stay as datums; service by local runners.
EXPERIMENTALTRL 3: lattice-walking robots are indoor prototypes; crease-rail service robot unbuilt
Risks we are designing against
- R1Ice, grit and water in sockets preventing lock
- R2Liability for falls and for people below
- R3Rail continuity across transverse creases
- R4Scope creep toward general-purpose robots
What the adversarial review changed
ProblemSOLL-E precedent stretched
NowSOLL-E framed as architecture precedent (304.8 mm lattice, about 15,000 indoor actions; successor about 5.6 kg), not scale
ProblemStatics do not close at 25 kg / 15 kg
Now25–35 kg body, 5–8 kg at full swing, 15 kg only in double stance; ≥ 400 N·m joints; step 15–30 s (0.03–0.06 m/s)
ProblemSingle expanding pin cannot carry moment
Now2–3 point cone/petal lock, 20–30 mm lead-in, ≥ 0.6 kN·m, ≥ 3 kN fall-arrest; sockets bolted to ridge beam; drained; lock sensing
ProblemRails conflict with fold creases; accuracy only local
NowSegmented rail on one panel, broken at transverse creases; building coordinates from survey
ProblemReach cannot cover slopes
NowExterior tasks limited to ridge-adjacent; PV washing to brush crawler or people
Problem6–12 modules per robot unsupported
NowHuman-ferried; one robot per contiguous rail network
ProblemCable kit underspecified
NowMoved to month 12–24; anchors on fixed frames
ProblemFlyCart misquoted
NowDJI official: 65 kg/12 km dual, 85 kg dual max, 100 kg single; steps say up to 65 kg routinely
ProblemARMADAS precision number mischaracterised
NowRestated as simulation: lattice RMS error collapses toward part error
ProblemExterior operation unaddressed
NowFall-arrest trolley, weather, grit and ice tests
Problem12-month scope overloaded; no human baseline
Nowv0 on 7.28 m bench rail + 3.64 m shell rail, 2 tasks x 200 trials; BOM 40–80k + labour 0.3–0.5M; human baseline measured first
Verdict: fix. The full log is on Evidence.
Every number on this page, with its basis and source.
| Quantity | Value | Basis | Source |
|---|---|---|---|
| SOLL-E: 5-DoF inchworm, about 15,000 indoor actions on a 304.8 mm voxel lattice (architecture precedent, not scale) | 5 DoF | measured | ntrs.nasa.gov/citations/20230005194 |
| ARMADAS foot petal capture of in-plane misalignment | 3.0 cm | measured | ntrs.nasa.gov/citations/20230005194 |
| SOLL-E successor robot mass | about 5.6 kg | reported | IEEE 11068540 |
| Lattice RMS error collapses toward part error as plate thickness grows (simulation) | qualitative | modelled | ntrs.nasa.gov/citations/20240000666 |
| Foot moment at the locked foot, 25 kg body + 15 kg at 0.91 m (rejected spec) | about 244 static, 370–490 with 1.5–2x factors N·m | calculated | 9.81 x (25 × 0.45 + 15 × 0.91) |
| Peak joint torque, 30–35 kg body + 8 kg at full swing | about 350–390 N·m | calculated | 9.81 x (m_body x 0.455 + 8 × 0.91) x 1.7 |
| Travel speed | about 0.03–0.06 m/s | calculated | 0.91 m per 15–30 s step |
| Industrial cable robot payload (rigid industrial frames) | up to 130 kg | measured | cms.tecnalia.com/uploads/2020/01/Catalogue_Cable… |
| DJI FlyCart 100: dual battery / dual battery max / single battery max | 65 kg for 12 km / 85 kg / 100 kg (short range) kg / km | claimed | DJI FlyCart 100 announcement and specifications (manufacturer) |
| Henn-na Hotel robots retired | > 50% of 243 robots | measured | engadget.com/2019-01-16-japan-robot-hotel-fires-… |
| Autonomous completion rate of contracted tasks | ≥ 95 % | target | Anyground gate |
| Robots per site | 1 per contiguous rail network, human-ferried between modules robots | target | Anyground |
Components
| Part | Material | Spec |
|---|---|---|
| TSURU body | Carbon/aluminium links, BLDC joints | 25–35 kg; 5–8 kg payload at full swing, 15 kg in double stance; 5 DoF; ≥ 400 N·m peak joints (target) |
| Foot lock | 2–3 point tapered cone or petal lock | ≥ 0.6 kN·m moment, ≥ 3 kN fall-arrest, 20–30 mm lead-in, lock-confirm sensor |
| Socket | Stainless cup bolted through to ridge beam | drained, self-cleaning; 910 mm pitch |
| Tool heads | Latch driver, thermal camera, cartridge gripper | tool changer on foot |
| Fall-arrest trolley | Stainless trolley on rail | continuous along exterior rails |
| Cable kit (month 12-24) | 4 winches, Dyneema | indoor bays only; anchors on fixed frames, not folding panels |
Sources
- ntrs.nasa.gov/citations/20230005194
- ntrs.nasa.gov/citations/20240000666
- doi.org/10.1109/IROS55552.2023.10341479
- doi.org/10.1109/ICRA57147.2024.10609866
- cms.tecnalia.com/uploads/2020/01/Catalogue_Cable…
- doi.org/10.1002/rob.21839
- engadget.com/2019-01-16-japan-robot-hotel-fires-…
- wikihouse.cc/product
Images and video on this page are AI-generated concept visuals. The 3D model is a concept drawing: dimensions follow the dossier, and details such as link lengths and part masses are illustrative.