Egocentric Pruning and Trellis Training Data

Egocentric pruning and trellis training data is real-world, first-person video of a person pruning and training plants. The site states five catalog rows for this. Four sit under Greenhouse, Nursery & Garden Centre: Prune with shears, which reads "Cut stems at the node with shears"; Stake + clip plant, which reads "Stake and clip a stem for support"; Trellis training, which reads "Weave stems into a trellis"; and Trim + deadhead, which reads "Trim spent growth and deadhead." The fifth, Pruning, sits under Farm & Field Operations and reads "Cut back dead and overgrown branches with shears and loppers to shape the plant."

People search for this under several names: pruning robot training data, a plant pruning video dataset, egocentric staking and trellis training data or custom horticulture data collection. They describe the same thing, a record of how a person cuts, supports and guides a plant, seen from the person own point of view, with both hands, the tool and the plant in frame.

Pruning and trellis work is not the same as potting, general plant care or grounds maintenance. The site states separate pages for plant potting and nursery work, for plant care and gardening, and for landscaping and groundskeeping, so this page keeps to the five rows above and treats those as neighbours. It does not cover residential yard-work rows.

This page states only what the site already states. Where a detail is not in the catalog or the delivery terms, it is marked Not stated and left for the written spec.

The four greenhouse rows are all rated Medium and listed as Available on request. The environment examples read "Greenhouse / grow-bench set: pots, trays, plants, shears, stakes, crates," so the rows sit on a bench set that names shears and stakes among its items. The Farm & Field Operations row, Pruning, is also rated Medium and is listed as Active collection. Its environment examples read "Farm fieldwork, greenhouse, orchard, vineyard -- harvesting, planting, pruning, irrigation, equipment operation."

The verbs in the five rows are the verbs a spec can build on: cut a stem at the node, stake and clip a stem, weave stems into a trellis, trim spent growth, deadhead, and cut back dead and overgrown branches. A buyer who needs only one of these can ask for it alone, and one who needs a full pruning-and-support sequence can ask for several together. The catalog does not split any row into finer steps, so any split is a spec choice and not a stated task list.

The site does not state which plants, crops or vines are used, which shears, loppers, stakes, clips or trellis types are used, how many cuts or plants a session covers, or what a finished plant looks like. Those are Not stated and are set in the written spec before capture begins.

Each episode follows the Firsthand episode schema: ten streams synced under 2 ms, with head and wrist RGB, depth, 3D hand pose with per-joint visibility, contact events on grasp and release, and frame-accurate action segments. For pruning and training plants, hand pose tracks the fingers on a stem and on the tool, contact events mark when fingers or the tool meet a stem, a stake or a trellis, and action segments mark each step of the sequence.

The site also states 200 Hz 6-DoF IMUs on the Firsthand capture rig. The four greenhouse rows list the capture mode Multiview rig + handheld grippers, which the site describes as a staged working set recorded by a multiview rig while the operator performs the task with handheld grippers, so demonstrations map onto a robot end effector. The Farm & Field Operations row lists Phone (head-mounted) / multiview rig. Confirm the stream list in the spec before you plan a training pipeline around a specific sensor.

Object labels use the terms your spec sets. Plant species, tool types, support types and a pruning-outcome or task-completion label are Not stated. If your policy needs a label such as "stem cut at node" or "stem secured," define it in the spec and the accept criteria. This page makes no head or camera pose, mapping or SLAM claim as a deliverable.

Delivery formats are RLDS, LeRobot and WebDataset, plus HDF5, zarr and Rerun .rrd. The format table below sets each against what its official documentation states. Pruning and trellis episodes use the same schema as every other Firsthand collection, so a loader that reads another Firsthand dataset reads this one without changes.

Every batch is validated against signed accept criteria before it ships, with a reject log, and delivery is rolling, per validated batch. Rejected items are recaptured at our cost. Billing is per validated hour, and every collection ships under a perpetual, buyer-owned commercial license.

The site does not state a number of episodes, a batch size or a total hour count for any of the five rows, so those are quoted against your spec and not assumed here.

Pruning, staking and trellis work are hand work with living plants. A person finds a node, closes shears on a stem, clips a stem to a stake or weaves it through a trellis. A policy that is meant to learn this needs to see how the hands and the tool move and when they touch the stem, not only the finished state of the plant.

First-person video suits that need because the view follows the worker, so the tool tip, the stem and the hands stay in view. 3D hand pose and contact events add the finger-level detail that plain video leaves implicit, and action segments mark where each step starts and stops so a model can learn the order as well as the motion.

This page quotes no robot result, no recovery rate, no throughput figure and no benchmark. It also does not claim that any model has been trained on this data, and it makes no yield, plant-health, chemical, pesticide, fertilizer, safety or certification claim. The site states five catalog tasks and a capture and delivery method, and the value to a given policy is something a pilot can test against your own accept criteria.

For the four greenhouse rows, the site describes a staged working set recorded by a multiview rig while the operator performs the task with handheld grippers, so demonstrations map onto a robot end effector. That is a staged set rather than a claim of recording unstaged real-time work. The Farm & Field Operations row lists a head-mounted phone or multiview rig, and the site does not say more about where that row is recorded.

The site does not state whose bench or plot that is, how many sites a collection uses, or which regions or cities are covered for any of the five rows. Elsewhere it states that Firsthand does not capture at a customer own facility, so this page makes no claim of capture there. Environment, access and schedule are agreed in the spec.

Every episode links to a signed contributor release through a documented consent chain, and the consent record is part of the delivery. Faces, plates and screens are detected and irreversibly blurred at ingest, and a reviewer confirms the result before anything ships. The anonymization record is delivered with the data.

That is the whole of the privacy wording the site states. It does not state a rule for data stored on a device, so none is claimed here, and raise it in the spec if it matters to you. This page also makes no yield, plant-health, chemical, pesticide, fertilizer, safety or certification claim.

The four greenhouse rows are listed as Available on request, which the site defines as scoped and recruited to spec when a buyer asks for it, so recording is not already running for them. The Farm & Field Operations row is listed as Active collection. The site does not state a separate timeline for any of the five rows, so each is quoted against your spec before any capture begins.

The site publishes two per-batch speed figures. Where coverage is already deep, the first validated batch lands in a median of 48 hours from signed spec. Thin environments typically take three to four weeks, because vetting operators there is the bottleneck, and we quote that up front. The site does not state which of the two applies to pruning and trellis work, so the quote says.

On-demand top-ups run against the same spec and accept criteria after each training run. The site does not state a batch size or a later-batch turnaround for any of the five rows, so those are quoted against your spec.

You describe the pruning, staking or trellis sequence your policy needs. We turn that into a written spec covering tasks, environments, capture tier, annotation depth and accept criteria, then quote reach, timeline and price before any capture begins.

A Spec Pilot, a small and fast first batch, can prove the spec and accept criteria before you scale. Billing is per validated hour, a reject log ships with every batch, rejected items are recaptured at our cost, and every collection ships under a perpetual, buyer-owned commercial license.

Useful things to settle in the spec are the actions you want covered, the plant types, the tools and supports, any label you want for step success, and the formats you want delivered. Each of those is Not stated by the site today, which is why the spec is where they are written down.