Cultivation Loom
A fixture-first, serviceable farm cell for Anyground
Design proposition — 29 October 2026. Make the crop system legible to machinery before making the machinery more intelligent. A cultivation loom is a small production cell in which a seed plug, crop collar, root channel, tool dock and replaceable service cartridge share a measured interface. The platform changes its crop-specific fixtures and tools, not its entire robot or building. This is a proposed integration of demonstrated mechanisms and commercial component classes, not a claim that an autonomous farm has already been built.
The first domain is deliberately concrete: compact lettuce and selected herbs, with controlled planting positions and a bounded canopy. It is a pathway toward a genuinely transformable living environment: the room can accept another root cassette, another tool cartridge or another environmental skin while keeping the same service spine. It is not an all-crop robot, a miniature staple-food economy, or a promise that every intervention disappears.
1. What is actually new in the combination?
Automated sowing, harvesting and cleaning already exist. Green Automation describes a commercial sequence including germination, adjustable gutter spacing, harvest, cleaning and a dry buffer; therefore “a robot grows lettuce” is not the innovative proposition. Its documentation concerns much larger greenhouse installations and is supplier evidence, not an independent benchmark. F15 — commercial process
The proposed contribution is interoperability at the biological-service boundary. A crop carrier is simultaneously a mechanical fixture, a traceable batch object and a wet-interface object. A common datum is useful only if the machine also knows whether the carrier is drained, which cleaning state it has reached, and whether its crop fits the current motion envelope. Treat these properties as a transaction, not as decorative dashboard telemetry.
Five combinations make this worth prototyping:
- Plant nursery × flexible manufacturing fixture. A small plug carrier first supports dense propagation, then locates into a larger growth rack. Interchangeable collars change spacing and root support without changing the robot's coordinate system. The key invariant is the carrier datum, not a universally sized plant.
- Japanese production engineering × fluidic interlock. “Ready to move” means mechanically released, identified and sufficiently drained. This is poka-yoke through geometry and state, plus jidoka: stop the affected operation when its prerequisites are uncertain. It is not a Japanese visual style.
- Robot tool changer × crop-specific hygienic cartridge. One dock carries different end effectors, while removable wetted parts and blade/pad inserts have their own cleaning identity. Standardize the dry mounting contract; do not pretend every crop-contact surface can share a universal wash recipe.
- Industrial maintenance transaction × autonomous service experiment. A pump becomes a keyed, accessible cartridge with capture wells, isolation proof and a post-installation flow test. This sharply bounded operation is a credible first step toward service autonomy; arbitrary robotic repair is not.
- Digital crop passport × mass balance. A carrier record combines seed lot, images, estimated crop envelope, measured mass changes, irrigation state, cleaning history and exceptions. Contradictions become diagnostic evidence. More AI does not repair a missing physical measurement.
These are integration hypotheses, not patent-novelty assertions. The disconfirming question is whether this interface discipline measurably reduces interventions and replacement time enough to justify its extra fixtures and sensors.
2. The physical cell and its universal boundaries
The model proposes an approximately 3.8 × 2.8 m frame, 2.0 m overhead rail height and a roughly 0.9 m root-channel level. Four removable channels contain six crop positions each: 24 sites, not hundreds of invisible productive plants. The nominal sensitivity model uses 2.0 m² of active canopy. Dimensions are initial engineering inputs; structural calculations, cable routing, guarding and manufacturing tolerances remain to be designed.
The gantry is a true hierarchy: the X bridge carries a Z cross-slide, which carries a vertical Y lift. Translation of the bridge moves everything attached to it. The lift descends only at surveyed locations. A downward RGB-D camera sees the crop; protected electrical and data services travel above the wet interfaces. Light bars occupy the fixed canopy and must clear both carriage and crop envelope.
Common interfaces have separate contracts:
- Mechanical: three locating features, an asymmetric key, positive retention, an accessible handle and a declared payload/center-of-mass envelope.
- Fluid: identified inlet and return, isolated ports, a defined drain direction, measured retained volume and captured residual drops.
- Electrical: dry contacts above the wet dock, tool identity, power class and safe de-energized state.
- Information: batch identity, calibration versions, cleaning state, supported actions and fault evidence.
FarmBot demonstrates that an open Cartesian platform can combine a multi-utility mount with interchangeable tools. Its software also makes tool locations explicit. These are useful implementation precedents, not proof that a small garden mount is suitable for wet hygienic harvesting. F01 — open tool interface, F02 — taught tool locations
3. From seed to a verified crop position
The fixed seed station accepts a keyed, lot-identified cassette. A replaceable pocket disk meters pelleted lettuce seed into a short chute over a plug carrier. Camera inspection labels every position empty, single, multiple or unknown. Neither an attractive image nor a successful motor step proves that a seed was placed. Rejecting an uncertain plug is cheaper than occupying a growth site for weeks with no plant.
The first trial uses a single seed format and a single substrate. The pocket insert is crop-specific. Loose irregular herb seed may need a different disk, vacuum needle or batch sowing strategy. A parked vacuum-seeder cartridge demonstrates that the common dock can accept another mechanism, but the animation does not claim an automatic tool swap occurred.
Germination is a real separate process: humidity, temperature, irrigation, darkness/light requirements and emergence time belong to the recipe. The model's transparent nursery lid exposes the carrier; it is not a validated climate enclosure. A nursery water circuit remains separate from mature-crop recirculation. Initially, a person transfers accepted young-plant carriers into the growth fixture. Robotic transfer becomes a later test of handle capture and drain-safe movement, rather than a hidden assumption.
This is an important Japanese precedent: Ohara and colleagues describe a simple conveyor robot and a bounded cultivation experiment. Their apparatus was revised to stop nutrient flow and drain before tray transfer after spills were observed. The lesson is not that their entire lifecycle was labor-free; it is that changing the crop-handling process can solve a robotics problem more effectively than a more elaborate arm. F03 — apparatus and transfer method
4. Root cassettes, water and resource circulation
Each opaque root cassette has a removable lid, replaceable plant collars, an upstream inlet, supported shallow channel and lower gravity return. Its slope is an adjustable design parameter to test with mature roots, not a rendered decorative angle. The exposed root bundles explain why an empty-channel flow test is insufficient: growing roots change hydraulic resistance and can bridge a return.
A low reservoir feeds the pump, filter and individually valved rows. The return header slopes toward a removable screen before the reservoir. Keep enough reserve for pump shutdown and return drainage, with a high-level trip independent of the nominal controller. A backup pumping strategy and a documented root-survival window matter more than a green status LED. NFT guidance supports this architecture while emphasizing appropriate head, channel support, monitoring and backup provision. F12 — NFT engineering guidance
Separate peristaltic channels meter nutrient concentrates and pH correction into a mixed reservoir region. Never mix incompatible concentrates together upstream of dilution. A dose is permitted only inside a configured mass/time budget and after sensor plausibility checks. It is followed by mixing delay and response verification. The controller records commanded and measured amounts; a worn tube or air bubble can make them diverge. Commercial embedded dosing hardware exists, but tubing life and compatibility remain measured maintenance tasks. F13 — dosing component class
EC is not a complete nutrient assay. A plausible bulk conductivity can coexist with individual-ion imbalance. Log replacement/purge water and nutrient inputs, take independent samples, and establish a crop-specific replenishment protocol. Level, leak, pH, EC, temperature and tray mass complement rather than replace one another. Dissolved oxygen is an additional trial measurement even though a probe is not depicted for every channel.
Recirculation can also circulate problems. A diseased row should isolate into a contained drain path, not be automatically flushed into healthy rows. Thermal or UV treatment would need dose, exposure, transmission/temperature, residence-time and bypass monitoring. Historical heat-treatment research includes heat recovery; it does not justify claiming that every return loop is sterile. F09 — treatment apparatus
Condensate is a third circuit. Transpired vapor passes through room air and a condensing coil into an untreated buffer. Coils, pans and air can introduce contamination or corrosion products. Reuse requires an appropriate treatment and water-quality release decision. No direct condensate-to-nutrient, wash-to-nutrient or potable cross-connection is shown. Recovery is useful, but a coil is not free water: cooling and dehumidification have a measured electrical and thermal burden.
5. How the robot really harvests
A known crop collar reduces the search space. The camera estimates the stem and canopy envelope; a deterministic controller chooses a collision-checked approach from surveyed fixture coordinates. Descent is vertical at a known site. Compliant fingers retain the head before a short guarded push cutter crosses the stem. Retention, severing and load transfer require separate confirmations. After cutting, lift vertically above the canopy before traversing to the clean output tray.
The compound cartridge in the model has carrier-handle fingers and a removable cutter, with distinct crop pads. It stays attached throughout the illustrative sequence. Seeder and soft-gripper spares remain in a physical magazine. A true automatic exchange would require a separate engage–lock–verify–retract trajectory and utilities test. Recent self-aligning tool-changer research suggests how lead-ins can make docking less sensitive to positioning error, but its limited tests are not a wear-life guarantee. Industrial positive-retention and lock-sensing practice supplies another boundary worth preserving. F07 — geometric alignment, F08 — industrial locking reference
Harvest literature supports a real prototype without supporting magic autonomy. A linked holding-and-cutting mechanism was evaluated in ASABE work. A 2026 in-situ harvester reports a 94% success rate in its particular tests; its push/pull cutting comparison makes support geometry an interesting design variable, not a universal force specification. F04 — cut/hold mechanism, F05 — in-situ mechanism
A particularly useful recent counterexample is a 15 October 2026 preprint: 94.4% reaching became only 86.1% complete harvest across 144 attempts. Reaching is not harvesting. Anyground should log perception, reach, capture, cut, retention, delivery and quality separately; uncertainty should not trigger repeated blind cutting. This preprint is promising emerging evidence, not established field reliability. F06 — full experiment
The output station independently confirms a mass increase and release image. Packaging, chilling, rejected leaves and the time spent cleaning a tray are real operations. A head that reaches the bin but is bruised, contaminated or unmarketable does not count as a successful food-production outcome.
6. Cleaning, quarantine and one repair that can be automated
The dirty front corridor has a wash trough, debris basket, waste tank and separate lidded quarantine tray. First remove roots and substrate; then wash, perform a validated sanitation process where required, rinse if applicable and dry. A completed wash timer is not a cleanliness measurement. Food-contact qualification, microbial validation, chemical compatibility and worker protection require appropriate specialist review before edible operation. The model deliberately specifies no universal sanitizer recipe.
Clean carriers return only after a recorded release. A dirty tool cannot approach clean seed stock simply because its tool ID is correct. Blade inserts and soft pads should be removable so inaccessible crevices do not become permanent cleaning problems. Commercial gutter systems already include cleaning and dry-buffer stages; small-cell automation must not omit them to make its autonomy claim appear better. F15 — cleaning and buffering
The service experiment is precise: one accessible pump cartridge with a vertical handle, two valved couplings and dry electrical connection. Stop the pump, isolate both sides, verify electrical/pressure state, capture a controlled drain, grip the handle, release retention and lift straight up. The animated module rises 0.32 m with its gripper; it does not fly independently. A second spare pocket is visible. Completing exchange still requires parking the old module, fetching and installing the spare, confirming retention/identity, leak testing, priming and measured return flow.
Commercial non-spill couplings make this credible today. CPC specifies finite residual spillage for its NS4 class: a small residual still demands capture and testing. “Non-spill” does not mean aseptic, zero dead volume or fertilizer compatibility. The precise coupling must be selected with its manufacturer; the drawing is not a CAD reproduction of that product. F14 — fluid connector specifications
A bounded diagnostic agent may compare observations, retrieve the cartridge procedure and prepare a service job. It does not receive unrestricted motor, valve or dosing authority. A local state machine enforces preconditions and an independent stop. Missing evidence escalates; it does not become an invented successful repair receipt.
7. Quantitative model: what this small cell can and cannot do
All following values are initial assumptions to expose trade-offs, not crop-trial results. At 2.0 m², 12 sites/m², 0.15 kg harvested head, a 35-day cycle and 85% saleable fraction:
fresh yield = area × site density × head mass × saleable fraction / cycle
This gives approximately 0.087 kg/day averaged across a steady staggered campaign. A 24-site prototype is an engineering test bed, not a proven economical household farm. At an explicitly illustrative food-energy assumption of 15 kcal/100 g, that is only about 13 kcal/day. Fresh greens can improve variety and micronutrient availability while contributing almost no staple calories; no claim of nutritional independence follows.
With assumed DLI 16 mol/m²/day and fixture efficacy 2.8 μmol/J:
LED kWh/day = canopy m² × DLI / (3.6 × efficacy)
The result is 3.17 kWh/day before HVAC, driver/optical losses, pumps, fans, controls or nursery lighting. A daylight greenhouse changes the balance but introduces solar heat and seasonal variability. At hypothetical transpiration 3 L/m²/day, 80% condensate recovery and 1 L/day purge, calculated make-up is 2.2 L/day, before wash, nursery, harvest water, leakage and treatment reject. Associated latent load at 0.68 kWh thermal/L is 4.08 kWh thermal/day; it is heat bookkeeping, not a free energy source or a direct electrical-consumption prediction.
For four channels totaling 1.2 L/min and 1.8 m head, ideal hydraulic demand is ρgQH; at assumed 20% pump efficiency, electrical input would be approximately 1.77 W. A real small pump may draw much more because its operating curve, minimum size and losses dominate. Measure installed head/flow and input power rather than buying from this idealized figure.
For a single 1.6 m traverse with 0.15 m/s maximum speed and 0.3 m/s² acceleration, a trapezoidal move takes roughly 11.2 seconds, before tool action, vision, vertical moves or settling. Harvest throughput must be end-to-end. At the preprint's illustrative 86.1% complete-harvest rate, 24 attempts would average about 3.3 exceptions if that rate transferred—which it may not. Remote intervention minutes can overwhelm motion-cycle savings.
EDEN ISS demonstrated real diverse-crop production in an extreme environment, while its crew-time research explicitly counted local and remote support. These are reasons to build an honest operations ledger, not to borrow Antarctic yield or labor as a forecast for this cell. F10 — crop campaign, F11 — workload study
8. Prototype gates, alternatives and adaptation
Build one transparent water-only row first. Verify drainage, leak containment, sensor disagreement and root-shaped blockage surrogates. Then add a guarded Cartesian rig, dummy crop collars and the cut/hold tool. Only after supervised motion and sanitation review should live crops enter. Run three complete crop campaigns before describing routine unattended intervals; log every onsite and remote minute, consumable, discarded batch and failed cycle. Proposed thresholds and cost allowances are in prototype.json, and may be revised when evidence contradicts them.
Compare at least three approaches with the same labor ledger: fixed fixture + gantry, moving gutters + stationary harvester, and human tending + inexpensive monitoring. A moving-gutter factory may win on throughput; a human may beat the small robot on total cost and irregular crops. The gantry wins only if a flexible small footprint and shared services justify it. Do not add a robot where gravity and a fixture solve the problem more reliably.
Herbs require different seeding density, pad shape and cut strategy; repeated cuts need regrowth and blade-cleaning evidence. Fruiting crops are not a software recipe switch. Tomatoes or cucumbers require larger supported root volume, trellis load paths, pollination, pruning, fruit detection and distinct end effectors. They may share the dry dock, service cartridges and crop passport while using a substrate/drip root cassette and a separate tall work envelope. That is meaningful universal modularity: common interfaces with honest domain-specific mechanisms.
In arid climates, compare a shaded daylight enclosure against a sealed lit cell using actual weather and water quality. In humid climates, dehumidification capacity and condensation hygiene dominate. Cold climates add insulation, frost-safe plumbing and heating/recovery integration. A room-scale platform should accept those climate skins without changing the crop datum or service transaction—but each adaptation needs a new energy and safety case.
The ambition is not a cautious catalog of existing devices. It is an inspectable, experimentally falsifiable architecture: make biological work fixture-compatible, make wet systems transaction-safe, and make autonomy measurable all the way through resupply, cleaning and failure. That combination can be constructed now. Whether it deserves to scale is exactly what the first prototype must reveal.
Next experiment: a deployable spacing cassette, not a folding greenhouse ornament
An especially promising connection to Anyground's origami platform is a dry-side pantograph nursery cassette. Closely spaced seed collars ride on a one-degree-of-freedom linkage above a removable wet liner. Pulling one indexed actuator increases collar spacing as the canopy grows; hard stops, not camera inference, establish the final crop locations. This combines deployable mechanisms, nursery density and manufacturing fixtures. Keep joints outside the nutrient stream so an interesting folding mechanism does not create an uncleanable root trap. The root liner must tolerate the change without pinching roots or forming stagnant pockets.
The first version should be a manually indexed, full-size six-collar strip—not a motorized room. Measure collar position at three indexed spacings, linkage force under wet mature-plant weight, pinch access, root disturbance and cleaning time. A proposed go/no-go is repeatability within 2 mm at the harvest index and no greater root damage or intervention time than the fixed-spacing control. Those are design targets, not established agronomic tolerances. A sliding comb with positive detents may outperform a foldable linkage; test both before choosing the more futuristic-looking mechanism.
Another non-obvious extension is to use the carrier itself as a calibration transfer artifact. Its underside can contain a protected reference patch and known geometric fiducials. Every passage through a clean inspection station checks camera scale and lighting drift, while a separate load-cell reference checks mass measurement. This imports metrology discipline from manufacturing into crop automation. It is useful only if the reference stays clean and its drift is independently checked; an ornate “digital twin” without this feedback would provide less trustworthy information.
Neither extension is claimed as demonstrated by the present scene. They are inexpensive, discriminating experiments that connect the farm to the transformable platform without substituting architectural spectacle for a working crop mechanism.