Short answer
Greenhouse and nursery demonstrations show potting, transplanting, spacing, pruning and tray moving with living plants that vary by species, age and season. Some operators may hold training or monitoring video, but plant diversity and gentle handling are likely to need deliberate capture across a growing cycle. Video gives no force or robot action data, and one crop does not represent another.
The buyer's problem
A greenhouse is semi-structured. Benches, trays, irrigation lines and aisles are laid out, and the objects on them are alive. Two seedlings of the same species look different; a mature plant has a branching shape that no CAD model describes; leaves tear, stems snap, and soil spills. The work is also seasonal, so the scene in March is not the scene in August.
A buyer who trains a perception or manipulation model on a single week of one crop risks a model that handles only that week. The page explains how to specify variation and how to treat gentle, delicate handling as a first-class data requirement.
Task anatomy
Propagation and potting
- Fill pots or cells with growing medium; level and firm it.
- Take a seedling or cutting by the right part (root ball or stem base), make a hole, place and firm.
- Water in, label, and move to a tray.
Handling and spacing
- Load trays onto a cart or conveyor; push to a bench.
- Space pots out as plants grow, or consolidate them.
- Rotate, turn or sort by size and quality.
Crop care
- Prune, pinch, tie or trellis.
- Remove dead leaves and flowers; check for pests.
- Harvest fruit or cut flowers; stage in crates.
Transport
- Move loaded carts between benches; see pushing carts and trolleys.
Housekeeping
- Clear spilled soil, empty trays, wash benches.
Failure modes are crushed stems, over-firmed soil, damaged roots, wrong plant to wrong pot, dropped trays, mislabeled crops and wet floors. Because plants are the product, damage is a loss the grower cares about.
Where public task vocabulary helps
The O*NET database includes Task Statements and Detailed Work Activities (DWAs), and a file mapping tasks to DWAs (O*NET Resource Center). Agricultural and horticultural occupations are described there at a coarse level. They help you name steps consistently across suppliers, and they do not describe plant-specific technique. Build your own crop-specific labels on top.
For adjacent produce work, see produce packing and grading, which concerns harvested goods rather than living plants.
Variation a buyer should specify
| Dimension | Why it matters | How to record |
|---|---|---|
| Species and cultivar | Shape, stiffness and fragility differ | Species and cultivar ids per clip |
| Growth stage | Seedling vs mature handling is different work | Stage label and approximate age |
| Season and lighting | Light, glare and condensation change appearance | Date, time and sky condition |
| Pot and tray type | Geometry, weight and grip points | Container model id |
| Medium and moisture | Soil weight and stickiness | Moisture note per session |
| Greenhouse type | Glass, polyethylene, shade cloth change light | Structure and covering note |
| Worker technique | Different growers pot differently | Worker id (pseudonymous) |
A collection spread over weeks of a growing cycle is likely to be more useful than one long day. Plan accordingly with the demonstration coverage plan, and expect rare events (pest outbreak, wilting, storm damage) to need deliberate attention; see rare events in demonstrations.
Sourcing
Existing footage. Some growers use cameras for security, climate monitoring or automated phenotyping. Check whether monitoring cameras are fixed on the plant canopy rather than hands. Training videos for seasonal staff may exist, and some operators record time-lapse of crops. These can help perception and plant-state recognition, though they are not hand-level demonstrations. Sensor logs such as climate and irrigation data appear in catalog guides like sensor and IoT data; they give context, not motion.
Commissioned recording. Operators can host sessions on real crops. You define species, stages, tasks and views, and ask for sessions at several points in the growing season. Ask whether workers are seasonal or contract labor, since that affects who can consent and for how long. Where workers are not employees, ask whether consent and rights run through the grower or the labor contractor, and confirm in advance. This is not legal advice. See worker consent and identifiable people.
Describe the crops and tasks you care about: tell SourceX the plant-handling data you need and it will investigate whether US businesses hold suitable material and discuss whether commissioned recording is feasible.
Illustrative package
# Illustrative only - synthetic values, not an available dataset
package: nursery_potting_pilot
site: nursery_site_015, covered bench area
crops: [crop_A herbaceous seedlings, crop_B woody cuttings, crop_C trailing plant]
sessions: 4 sessions spread across one growing cycle
workers: 6 (grower_J01..J06), written consent via employer and individually
tasks: [pot_fill, transplant, tray_load, spacing, prune_pinch]
views: head_mounted; bench_overhead; side_close_up on hands and root ball; cart_view
labels: step, plant_id, growth_stage, container_id, handling_part (root_ball/stem/leaf)
outcome: intact / damaged_leaf / broken_stem / root_exposed / replanted
metadata: date, light condition, medium moisture note
excluded: grip force, soil firmness, robot actions
rights: training and evaluation; no redistribution
Quality and acceptance pointers
- Plant visibility. Check that the root ball, stem and hand contact point appear in the close view, since a canopy-level view hides most of the handling.
- Damage labeling. A reviewer with horticultural knowledge should label damage; the person who did the task may not notice. Keep damaged outcomes in the dataset.
- Season spread. Confirm sessions cover more than one date range, and record the crop stages at each.
- Appearance range. Wet leaves, condensation on glass, shadows and humid camera lenses all occur. Samples should include them.
- Evaluation split. Hold out a cultivar, a season or a nursery. A random frame split would leak plants across train and test.
What this data cannot support
- Plant mechanics. Stiffness, break strength and the force that bruises a stem are not visible. A robot needs trials on real plants, ideally with grip sensing.
- Biological outcomes. Whether a transplant survives is known weeks later and requires follow-up records, not clips.
- Robot actions. Human video has no joint states or gripper commands. Plant handling may need low-force, compliant control, which buyers typically develop with robot data on hardware.
- Cross-crop transfer. A potting dataset for one herb says little about orchids or trees.
- Agronomic decisions. When to prune or water is an expert judgment recorded in documents and conversation. Fragile item handling covers handling labels, not the horticultural rationale.
Questions to ask suppliers
- Which species, cultivars and growth stages appear, and over what dates?
- Is the footage from monitoring cameras, training videos or purpose-built sessions?
- Can hand-to-plant contact be seen in a meaningful share of clips?
- Are damaged plants and failed attempts kept and labeled by someone with horticultural expertise?
- How are plant, pot and tray ids recorded?
- Who employs the workers, and who can grant consent for model training?
- Were visitors, customers or other businesses ever in view?
- What lighting and humidity conditions are covered, including condensation?
- Do sessions span more than one season or crop cycle?
- Can a species, season or site be held out for evaluation?
Discuss your requirements
Discuss your humanoid training-data requirements with SourceX by describing the plants, tasks, views and permitted uses. SourceX reviews the requirement and follows up about potential sources; a request does not guarantee a match, and nothing is contracted until a supplier agrees. For wider context, see the workplace task demonstration hub, the humanoid training data guide, and deformable object handling.
Describe the workplace demonstration data your robotics project needs
Sources
- ONET Database (ONET 29.1 per search snapshot): Task Statements and Detailed Work Activities — task statements and Detailed Work Activities as a task vocabulary.