Stereo Crop Harvesting Data

It is first-person video of a person harvesting produce by hand or with shears, recorded with two calibrated cameras, with hand pose and sensor streams alongside. People reach it through several phrasings: harvesting robot training data, stereo hand harvesting video, egocentric produce picking with shears dataset, or custom agriculture data collection at scale. They describe one request from different angles. Stereo describes the capture, and MCAP describes the container. They are separate choices.

The site does not state that stereo is already running on these rows. The catalog lists a head-mounted phone or a multiview rig as the capture setup for the farm rows, so a stereo pair is a capture choice agreed in the skill spec and confirmed on a Spec Pilot. The data is human capture by paid, consented contributors, not robot or teleoperation data, so using it for a specific robot still needs your own retargeting step. Firsthand is not affiliated with Foxglove or the MCAP maintainers.

Four, listed in lib/commercial-catalog-data.ts, in the Agriculture & Grounds sector. Three sit in Farm & Field Operations, all rated Easy and all Active collection: "Harvest and sort produce using pruning shears and field crates", "Harvesting" and "Agriculture Hand Harvesting". One sits in Greenhouse, Nursery & Garden Centre: "Harvest ripe produce", rated Medium and listed as Available on request.

The site defines Active collection as "Recording in this environment is running now, so a spec can join an existing capture stream." That is a statement about the catalog row, not a promise of a delivery date, and Available on request means a spec starts the work rather than joining a running stream. The site states no delivered hours and no delivered stereo footage for any of the four rows, so this page claims none. The first table below reproduces the rows as the catalog states them.

This page is the stereo and MCAP treatment of hand harvesting: picking ripe produce, cutting it with pruning shears, placing it in field crates and sorting it in the field. The egocentric farm and field task page covers all 21 Farm & Field Operations rows as first-person video, including these three harvesting rows, and stereo is not its subject. It also lists a separate crate loading and hauling row, which this page does not claim.

The stereo produce sorting and packing page covers five rows about sorting, grading, weighing and packing produce in a packhouse or on a grow bench, all Easy and Available on request. That is work after harvest. The stereo landscaping and groundskeeping page covers one row, mowing, trimming and planting on commercial property, rated Medium and Active collection. The second table below sets the pages side by side with each status as the site lists it.

The catalog lists "Phone (head-mounted) / multiview rig" for the three farm rows and "Multiview rig + handheld grippers" for the greenhouse row. The site does not state a stereo pair for any of them. Stereo applies through the site’s general stereo offer: a stereo pair is a capture choice agreed in the skill spec and confirmed on a Spec Pilot. The site also states that Firsthand does not capture at a customer’s own facility, so the setting is agreed in the spec.

The site describes two stereo setups and does not say which one an order for these rows uses. The hard-environments stack is a global-shutter 4K/60 head camera with a 120° lens, two 1080p/60 wrist cameras, a depth sensor at 848×480 and 30 fps, and 200 Hz 6-DoF IMUs. The off-the-shelf stereo head rig, dataset 06, is dual fisheye cameras at 1080p per eye, a 62 mm baseline, per-unit calibration, a 9-axis IMU and hand-tracking output. For the greenhouse row, the handheld grippers named in the catalog are part of the stated setup; whether a stereo pair is worn, mounted or combined with them is Not stated.

Harvesting is about where a hand or a blade meets a stem: how a fruit is held, where the shears sit against the plant, how a crate is approached. A stereo pair recovers depth from the calibrated frame pair rather than from a standalone depth sensor, so the geometry sits in the same capture that shows the hands working. The site gives no depth accuracy figure for these rows, so what a buyer does with that depth is their pipeline’s question.

A stereo pair adds depth, not a verdict. The site states no ripeness, quality or success label for a pick, and none is claimed. The egocentric farm page already supplies the first-person framing; stereo is the addition for a buyer who wants the depth and the MCAP file.

The rows give a natural cut: pick ripe fruit or vegetables by hand or with shears, cut produce from the plant, check quality, sort, and place in crates. The greenhouse row adds "twist or snip" as the two ways of picking. A spec that says "harvest the crop" gives one long action segment. A spec that names the steps gives a sequence you can slice by step, and the site does not set an episode length or say how many picks make up an episode.

Whether picking and cutting are one segment or two, and whether placing in a crate and sorting are one segment or two, is a spec decision, not a default. Contact events can mark when a hand meets an item, where the spec asks for them, and the spec decides what counts as the object. The site does not describe contact events as force or pressure.

A leaf, a thumb and a pair of shears can all sit between the lens and the work. The site states that plants occlude the fruit and the hands, and that the hands occlude the work. That is why each hand joint carries a visibility flag, why contact events mark when the hand meets the item, and why instance masks are recorded per frame. It also states that reach and grasp frames are where blur and occlusion show up, and the head camera in the hard-environments stack is global-shutter, which avoids rolling-shutter skew.

The site gives no figure for how often foliage hides the work on these jobs, so a spec can ask for wrist views or a second angle and check it on a Spec Pilot. Tools are named only as the rows list them: pruning shears, shears and field crates. The site states no tool-pose, force or tactile stream for these rows.

Not much, and this page says so rather than implying a fix. The farm and field page states that the site lists no lighting or weather condition for this environment. The site does state that glare-heavy and low-texture scenes make single-frame depth worse, and that depth degrades on wet and reflective surfaces, with confidence maps masking low-confidence regions. A sunlit leaf, a wet fruit and a plastic crate are plausible cases, but the site does not say how often they occur or how depth behaves on them.

Where a program needs a specific time of day, shade, covered growing space or weather window, that goes into the spec and is checked on a Spec Pilot. Season, region, crop variety and growing system are Not stated for these rows. An episode is rejected when more than 35 percent of depth pixels in the manipulation region are invalid, and for this job the spec decides what the manipulation region is.

Blade handling, protective equipment, field safety, food safety and certification are Not stated for these rows, and this page makes no such claim. Which tool is in the hand and how it is used is set in the skill spec using only the tools the rows name.

For every delivered episode the site states that faces and screens are blurred irreversibly, that a reviewer confirms no un-consented person is identifiable, and that footage that cannot be de-identified is discarded. The anonymization record ships with the data, an undocumented bystander is a stated reject criterion, and every episode links to a signed contributor release. Who may be in frame and what the footage may show are agreed in the spec. This is not legal advice, and the executed agreement controls.

MCAP is an open source container file format for multimodal log data, supporting multiple channels of timestamped pre-serialized data. It is serialization-agnostic, so messages can be Protobuf, DDS (CDR), ROS, JSON and more. Its design is row-oriented and append-only, it stores schemas alongside the data, and it offers an optional index plus optional LZ4 or Zstandard chunk compression.

Reader and writer libraries exist for C++, Go, Python, Rust, Swift and TypeScript, and the mcap CLI can inspect, merge and split files. The foxglove/mcap repository is MIT-licensed; that is the MCAP project licence, not the licence of a Firsthand delivery. The MCAP hub page and the glossary entry cover the format in full.

Firsthand offers MCAP delivery, and dataset 06 is stated as shipping as MCAP. For these rows the site states hand pose, a stereo-derived depth stream with confidence maps, IMU, contact events, action segments, per-frame instance masks and one task instruction per episode elsewhere. What it does not state is which of those are channels in an MCAP file, how they are named, what encoding they use or at what rate.

No topic name, message schema or rate is stated on this page, and none should be assumed. Name the streams that matter in the skill spec and confirm them on a Spec Pilot batch. The fourth table below sets each stream beside what the site states about it.

Foliage and hand occlusion is the first risk, under the visibility flag above. Weak depth on glare-heavy or wet surfaces is the second. Calibration is measured at the start and end of every session under a 0.28 px RMS gate, and an episode is rejected for a sync breach beyond the ceiling for more than three consecutive frames, for the manipulated object out of frame for more than 15 percent of the episode, for an undocumented bystander or for a mismatch against the named skill spec.

A reject log ships with every batch, and rejected items are recaptured at the provider’s cost. The fifth table below lists the failure modes beside what the site states.

The October 2026 off-the-shelf catalog lists ~324,800 hours across 18 datasets, of which stereo is ~64,800 hours across 9 datasets. The site names no harvesting dataset among them, so this work is custom collection, not an off-the-shelf purchase. The off-the-shelf stereo and hand-tracking video can still show the format before you write a spec.

It is custom collection at scale through a network of thousands of capture providers, for a consistent, recurring data flow to labs. Where coverage is deep, the median time from signed spec to first validated batch is a median of 48 hours; where it is thin, turnaround runs three to four weeks. The three farm rows are Active collection and the greenhouse row is Available on request, but the site does not say which speed applies to which, so the timeline is quoted against your spec. Collections are billed per validated hour, and rejected episodes and recaptures are never billed.

Describe the steps and where each starts and ends: picking, cutting with shears, checking, sorting and placing in crates. Say whether an episode covers one pick, one crate or a longer stretch of work, what counts as the object and what counts as a finished step, the privacy rule for the space, and name the variation your pipeline is missing. Then name the streams you want in the file, the topic naming and message encoding your reader expects, whether calibration ships inside the file or beside it, and whether MCAP is the only delivery or sits beside RLDS, LeRobot or WebDataset.

Most programs begin with a Spec Pilot, a small first batch that proves the spec and accept criteria on real footage before you scale. The sixth table below shows what each spec element decides.