Egocentric Herb Bunching and Banding Data

Egocentric herb bunching and banding data is recorded evidence of one person gathering cut herbs into a bunch and banding it, collected by a paid, consented contributor who performs the task written in a skill spec. Each attempt is delivered as an episode with synchronized streams and labeled action segments. It is human capture. It is not robot or teleoperation data, and it is not a fixed corpus you download.

The site states the task in one row of its commercial task catalog. In the Greenhouse, Nursery & Garden Centre environment, the row "Bunch + band herbs" reads: "Bunch herbs and band them." It is a Medium row and its status is Available on request. The environment is described with the examples "Greenhouse / grow-bench set: pots, trays, plants, shears, stakes, crates". This page is built on that row and on the neighbouring greenhouse rows quoted below, and on nothing else the site says about herbs.

The word egocentric in the title needs a plain reading. Firsthand’s general offer is first-person capture, and the catalog lists a head-mounted phone for many rows. For this row the catalog lists something different: a multiview rig with handheld grippers. This page therefore describes the capture setup as the row states it, and treats first-person or stereo capture as a choice made in the skill spec, not as something the site promises here.

People reach this topic through herb bunching robot training data, egocentric herb banding video, a greenhouse produce handling dataset or custom horticulture data collection. They describe one request from different angles. The work is the same: write the task down, collect it on demand and receive validated episodes.

One row covers it. "Bunch + band herbs" sits under the Agriculture & Grounds sector in the Greenhouse, Nursery & Garden Centre environment. Its description is "Bunch herbs and band them." Its skill rating is Medium, its capture setup is "Multiview rig + handheld grippers" and its status is "Available on request". The table below reproduces the row, and the rows beside it that this page refers to, as the catalog states them.

On the site, "Available on request" means the work is scoped and recruited to spec when a buyer asks for it. It does not mean hours are sitting in a library. This page claims no delivered hours for the row, and no off-the-shelf herb bunching dataset exists; the work is custom collection. The status is shown here exactly as the catalog lists it, and it is the same for every row in this environment.

The site names no herb variety, no band type, no bunch size and no item count. The words that appear are herbs, band and bunch, and this page uses only those. Whether a bunch is a handful or a fixed number of stems, and what counts as a band, is not stated. Those choices belong in the spec, and a Spec Pilot is the place to settle them.

The catalog states one capture setup for this row: "Multiview rig + handheld grippers". The site explains it 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. That is the description this page uses, and it is the only capture description this page makes for the row.

The site does not state a head-mounted phone for this row, and it states no stereo pair for it. A calibrated stereo pair is offered generally, and a head-mounted stereo rig is the instrumented tier of the general capture stack. If a program wants either for herb bunching, it is agreed in the skill spec. The egocentric label in the title names the general first-person offer and the query people use; it is not a claim about this row’s rig.

The multiview setup matters for a task like this because a bunch is held in a small volume and every hand moves around it. Synchronized third-person angles show how the two hands share the work and where the band goes on. A rig that captures them with handheld grippers in the operator’s hands also lines up with the site’s stated reason for the setup: the demonstration maps onto a robot end effector. The site gives no camera count, resolution or frame rate specific to this row.

The site does not describe how hand pose applies when the operator is holding a gripper rather than bare hands. Hand pose with per-joint visibility is a general stream, and whether your program wants it, or wants the gripper state instead, is settled in the spec. Force and tactile readings are not among the listed streams.

Each attempt is delivered as one episode, and bunching needs a decision up front about what one attempt is. It could be a single gather, one finished banded bunch, a run of several bunches, or the stretch from loose stems to a bunch placed in a crate. The site does not set an episode length, a number of bunches per episode or a number of episodes, so the skill spec does.

A short episode is easier to check against a single accept criterion, such as one bunch banded and set down. A long episode shows the order of the work: how stems are picked up, how they are squared, when the band goes on and when the contributor moves to the next bunch. That is an illustration of how a spec can cut an episode, not a stated duration.

The cut matters because of one machine-checked rule. An episode is rejected if the manipulated object is out of frame for more than 15 percent of it. In a bunch the manipulated object is a handful of stems, and the contributor will turn and reach for the band. The spec should say what counts as the manipulated object at each step so the rule is applied the same way across contributors.

A second rule applies to depth: an episode is also rejected if more than 35 percent of depth pixels are invalid in the manipulation region. Thin stems and leaves are the kind of subject where depth can be patchy, and the site says depth degrades on wet and reflective surfaces and ships per-frame confidence maps. It does not say how depth behaves on herbs, so this page makes no accuracy claim for them.

Inside each episode, action segments divide the recording into start- and end-timestamped spans, each carrying start_ns, end_ns, verb and noun. For this task a segment can be one step as written in the skill spec, so a bunch becomes a run of labeled steps, such as gather, square and band, with a noun for the item, rather than one long clip. Other standard fields include hand_joints and hand_visible for 3D hand pose, contact_events, objects and instance_masks, stored at the depth resolution of 848×480 so they line up with the depth stream.

Hand pose is delivered with a per-joint visibility flag for occluded fingers. That flag is the signal that matters when fingers close around stems and a band. The site describes contact events as the moment the hand meets an object. It does not say which objects count in a bunching episode, so the spec settles what a contact event means for your task.

The site lists no label for the outcome of the bunch. There is no stated bunch-tightness, band-tension, stem-count or task-completion score, and no contact force reading. The catalog verbs describe what the contributor does, not a delivered measurement. A head trajectory and a map of the bench are likewise not among the listed deliverables. The site states 6-DoF IMU data at 200 Hz from head- and wrist-mounted sensors, and nothing more about motion.

Delivery is in the format your stack already reads: RLDS, LeRobot, WebDataset, HDF5, zarr or Rerun .rrd, or raw streams with metadata. Readable converter source ships alongside the data so the schema can be retargeted to your own loader. The official documentation for three of these formats is linked in the sources below. Collections are billed per validated hour, and the license is perpetual, irrevocable and worldwide, buyer-owned and non-exclusive by default, with exclusivity available as a separate term.

Hand occlusion is the main one. In a bunch the fingers, the stems and the band all occupy the same few centimetres, and the hand closes around the work. The site’s answer is hand pose with a per-joint visibility flag, so a model can tell a hidden finger from a missing one, and multiview angles that see the hands from more than one side. The spec decides which angles matter for your task.

The second is stems that look alike. A pile of cut herbs is many similar-looking objects, and a per-frame instance mask is only useful if the spec says what is one object. The site delivers per-frame instance masks where the spec calls for them, and it does not say how well they separate individual stems, so this page claims nothing about it.

The third is the object leaving the frame. A bunch is carried to a crate, turned over and set down, and the manipulated object can drop out of view. The 15 percent out-of-frame rule exists for that, and the episode is rejected rather than quietly accepted. Rejected items are recaptured at our cost, and only footage that clears the signed accept criteria is billed as a validated hour.

The fourth is sync and calibration. Streams are hardware-triggered onto one PTP clock domain and aligned under 2 ms, and the rig is calibrated twice per session, before capture and after. An episode is accepted only if RMS reprojection stays under 0.28 px. The site states fine motion and inspection checks as part of its quality process in general terms, and this page adds no herb-specific tolerance to them.

Bunching is a useful shape of task because it combines several hard parts of manipulation in one short sequence: gathering many deformable stems, holding them as one object, and fastening them with a second object. The site rates the row Medium, which is the same rating it gives to pruning, staking and transplanting in the same environment. The task is not simple, and this page does not claim that any current robot performs it.

First-person and multiview human video of the task sits at the base of the data pyramid the glossary explains, where volume is high and the gap to a robot body is widest. Human capture does not replace robot data, and this page does not claim that it does. The staged working set with handheld grippers is the site’s own way of narrowing that gap for rows like this one, because the demonstration is made with a tool that maps onto an end effector.

A custom collection lets you aim at the step your model is missing. If your pipeline already handles gathering and fails at the banding, or fails at the sequence, the spec can name that. This page does not claim that any public corpus lacks herb data, and it does not claim a training outcome or a model improvement from any dataset.

Every contributor signs a written release before capture, and each delivered episode links to that record through consent artefact IDs in the batch data card. At ingest, automated detection finds faces, plates and screens and blurs them irreversibly, and speech is scrubbed of names, addresses and card numbers while ambient audio is kept. A human reviewer confirms that no un-consented person is identifiable in any frame, and footage that cannot be de-identified is discarded. An undocumented bystander is also one of the machine-checked reject criteria. The anonymization record is delivered with the data.

The site names no grower, farm, nursery, region or city for this row, and it lists capture at your own facility as not offered. This page therefore makes no claim that episodes are recorded at a particular greenhouse, and none that they are recorded at yours. Capture is run to your skill spec using the settings and contributors Firsthand arranges. The site states nothing specific about workers or crops visible in the background of a greenhouse, so this page makes no claim beyond the general process.

On handling safety, the site states no sharps, shears or tool-safety protocol, no contributor safety protocol and no agricultural certification for this row, and this page does not describe one. If your program needs terms about the tools used or the conditions under which the work is done, raise them when you scope it. None of this is legal advice, and the executed agreement controls.

The table below sets this page against the neighbouring greenhouse, nursery and produce pages on this site. Each row states the page’s main subject and the catalog rows it draws on, with the status as the catalog lists it. No delivered hours are claimed for any of them.

The short version: the potting and seed tray pages cover filling pots and sowing and replanting seedlings, the pruning and trellis page covers cutting and supporting stems, the produce packing page covers packing and grading, the crop harvesting page covers picking ripe produce in stereo, and the plant care page covers the wider run of gardening tasks. This page treats the bunch and the band as the subject.

From signed spec to first validated batch, the site states a median of 48 hours where coverage is deep and three to four weeks where it is thin. It does not say whether the Greenhouse, Nursery & Garden Centre environment counts as deep or thin, so ask which timeline applies when you request the quote. Speed comes from a warm pipeline, not from cutting consent, review or anonymization.

The site also states program milestones: a pilot reaches its first delivery in two to three weeks, and a standing program reaches full rate in four to six weeks. Those assume capture coverage is already in place. After the first batch, on-demand top-ups run against the same accept criteria, so you can add herbs or settings without rewriting the spec. The site does not state a turnaround for those later batches, so plan them in your quote.

The site does not publish a price list, a minimum order or a batch size for this row, so those are set in your quote. Collection is fast and on demand in the sense that you do not need a standing program: a focused brief is a normal starting point.

Where this site does not state something, this page says "Not stated" rather than filling the gap. There is no herb variety, no band type, no bunch size or stem count, no task duration, no site count, no region or city, no episode count, no bunch-quality or completion label and no tool-safety guidance. Nothing here reports a robot result, a benchmark or a model improvement.

Counts of contributors, hours and tasks are set in your skill spec. A task outside the rows quoted on this page, such as a packing or labelling step the catalog does not name, is settled in a Spec Pilot rather than assumed.

Describe what one bunch is, what the band step looks like, where the finished bunch goes and the variation or failure cases your pipeline is missing. That becomes a written skill spec, quoted for reach, timeline and price before any capture begins, and most programs run a Spec Pilot first, a small, fast first batch that proves the spec and the accept criteria on real footage before volume scales. The scoping checklist below maps the parts of a skill spec onto a bunching request. A published sample recording is on the sample data outputs page if you want to see the streams before you write a spec.

Three choices matter most. Name the step and the item rather than a broad activity, because that is what lets each gather and band become one action segment. Decide whether you want the capture setup the row states, a multiview rig with handheld grippers, or a first-person or stereo rig agreed in the spec. And name where the episode starts and stops, since the site does not set an episode length.