Egocentric Barista and Espresso Data

It is real-world, first-person video of a person working a coffee bar: grinding beans, dosing and tamping, pulling espresso shots, steaming milk and pouring drinks at the bar. It is recorded from the operator viewpoint, with 3D hand pose, depth, contact events and frame-accurate action segments aligned on one hardware clock. The task is the one the site’s task catalog lists in a single row.

The data is human capture by paid, consented contributors, not robot or teleoperation data. This page is about barista robot training data in the sense of what a policy can learn from watching skilled bar work, and it makes no claim that any robot has been trained on it.

One. In the Cafe, Bar & Beverage Service environment of the Food & Beverage sector the catalog lists "Professional Barista Workflows": "Grind beans, dose and tamp, pull espresso shots, steam milk and pour drinks at the bar." It is rated Medium, it lists a head-mounted phone or a multiview rig as the capture setup, and its status is Available on request. The environment examples are a coffee shop, a juice bar, a cocktail bar and a hotel bar.

The site defines Available on request plainly: the work is scoped and recruited to spec when a buyer asks for it. This page therefore claims no completed collection, no delivered hours and no volume for this row. It names only the steps that appear in the row: grind beans, dose and tamp, pull espresso shots, steam milk and pour drinks, at the bar. The first table below reproduces the row as the catalog states it.

The egocentric beverage and coffee prep page treats drink making as a family of tasks: barista work, pour-over coffee, bar drinks, coffee machine operation and serving, across five environments. Barista work is one row among them, and it is the one that page calls the most complete beverage row.

This page narrows to that one row. It goes through the five barista steps in order, sets out how the capture works at a bar, what each episode carries and where the failure modes are, and leaves pour-over, bar drink building and machine operation to the beverage page. The comparison table below sets the pages side by side.

Grinding beans is the first step. It is a short, repeated hand motion at the start of a longer chain, and the data needs to show where it ends. Dosing and tamping follow: placing the ground coffee and pressing it down are small, fine-motor steps where the fingers and the pressing hand matter more than the wide scene. Pulling the espresso shot comes next, and the key signal is the order of the steps and the moment a hand starts and stops each one.

Steaming milk and pouring drinks close the sequence. Milk moves from one vessel to another and the drink is poured at the bar, so a spec decides whether the pour is a step of its own or the end of the steam. The site states no drink list, recipe or vessel size for the row, so which drinks are made and in what vessels is agreed in the spec rather than assumed.

Across all five steps, action segments carry the start and end of each one as the skill spec defines them. A spec can ask for the whole workflow in one episode or isolate a single step, and the segment boundaries are what let a policy learn the order.

A bar is a small workspace where the hands move between several pieces of equipment without the body moving much. A head-mounted camera keeps that whole reach in frame from the viewpoint a manipulation policy will see from, and a wrist-mounted camera is available alongside it for close-range detail on the grip and the tamp.

A multiview rig adds synchronized third-person angles when the spec asks for them. That is useful at a bar, where a hand can pass behind the equipment and out of the head view. Which angles a barista order uses is set in the spec.

The catalog lists one capture setup for the row: a head-mounted phone recording the first-person view of real work on site, with a multiview rig adding synchronized third-person angles when the spec asks for them. The site does not state a camera count, resolution or frame rate specific to barista work, so those are agreed in the written spec rather than assumed.

Streams are hardware-triggered onto one clock, synced under 2 ms. Depth is aligned to video, and distance-invalid pixels beyond the sensor reliable range are flagged rather than interpolated. The site does not state which cafe, coffee chain, machine or grinder a collection runs in or on, and it does not state capture at a customer own facility, so the environment, access and schedule are agreed in the spec.

Each episode follows the same episode schema as every Firsthand custom collection: ten streams synced under 2 ms, with head and wrist RGB, depth, 3D hand pose with per-joint visibility, contact and grasp events and frame-accurate action segments. The only language data is one task instruction per episode.

For barista work, hand pose tracks the operator through the grip, the press and the tilt. Contact events mark when the fingers make and break contact with a tool, a vessel or a control, and action segments mark the start and end of each step. Because the labels share one clock with the video and depth, a step boundary lines up with the exact frame where it happens.

Occlusion is the first risk. Equipment at the bar and the vessel in hand can hide the fingers that hold them. The per-joint visibility flag is the answer the site gives, contact events mark when a hand meets an object, and a spec can ask for wrist or multiview angles so a grip the head camera loses is still recorded.

Step order and step length are the second. The row chains five steps and the site does not say how long an episode is, so the spec decides whether one episode is a single drink, a single step or a run of drinks. A policy that is meant to learn the order needs the boundaries marked, which is what the action segments are for.

Depth is the third. Distance-invalid pixels are flagged rather than filled in, but how depth behaves around steam, liquid or reflective equipment is Not stated, so a spec should check it on a Spec Pilot. Bystanders are the fourth, since a cafe or a bar can put customers or colleagues in frame, and an undocumented bystander is a stated reject criterion.

Every batch is validated against signed accept criteria before it ships, with a reject log. The site states that an episode is rejected when the manipulated object is out of frame for more than 15 percent of it, when more than 35 percent of depth pixels in the manipulation region are invalid, or when an undocumented bystander appears.

Billing is per validated hour, and rejected items are recaptured at our cost. The site does not state an accuracy figure for barista step labels, so what counts as a pass for a step boundary is set in the accept criteria.

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. A coffee bar can put customers in frame, which is why that pass matters here.

Text on cups, labels and packaging is not listed as a blur category on the site, so this page does not claim it is blurred. If text on a cup or a bag of beans matters for your use, raise it in the spec so the handling is agreed before capture.

Drink recipes, vessel sizes, liquid volumes, fill-level labels, shot times, milk temperature, throughput, force or tactile data, a task-completion label and any food-safety or hygiene certification are not stated anywhere on the site as standard fields. No coffee or equipment brand is named for the row, and no customer is named.

If your policy needs any of these, they are defined in the written spec as annotation depth, and the quote covers them before capture begins. The page does not promise them by default, and it makes no food-safety claim.

The October 2026 off-the-shelf catalog lists ~324,800 hours across 18 datasets. None of them names barista or espresso work. One stereo set, S2 (~5,000 hours, Cambodia, 2, 4 and 6 camera rigs, stereo RGB, task labels and hand keypoints), lists cafés among its environments and "food and coffee prep" among its tasks. Whether S2 holds any of the barista steps, grinding, dosing and tamping, pulling shots, steaming milk or pouring at a bar, is Not stated. Another mono set, M2, lists food service among its commercial settings.

Barista and espresso collection to a spec is therefore custom collection. We run custom collection at scale through a network of thousands of capture providers, for a consistent, recurring data flow to labs. The site states no provider count or throughput for this row.

Delivery formats are RLDS, LeRobot and WebDataset, plus HDF5, zarr and Rerun .rrd. The first three are training-loader formats, HDF5 and zarr carry arrays for custom pipelines, and Rerun .rrd gives a synchronized playback file for review. The format table below sets each against what its official documentation states.

Converter source is readable, so you can retarget the schema to your own loader, and the episode schema page lists the fields. Delivery is rolling, per validated batch, and formats are confirmed when the work is scoped.

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 row is Available on request, so the scoped-to-spec timeline applies rather than the ready-now case, and the quote states it up front.

On-demand top-ups run against the same spec and accept criteria after each training run, so new barista data stays consistent with what you already hold. The site does not state a batch size or a later-batch turnaround, so those are quoted against your spec.

Describe the steps, the drinks, the equipment and the bar you expect to see. Say whether one episode is a single drink, a single step or a run of drinks, what counts as a reject, and which angles you want beyond the head camera. We turn that into a written spec covering tasks, environments, capture tier, annotation depth and accept criteria, then quote reach, timeline and price against that document before any capture begins.

Most programs begin with a Spec Pilot, a small first batch that proves the spec and the accept criteria on real footage before you scale. The collection runs through the same seven-stage end-to-end custom collection pipeline as every Firsthand program, under a perpetual, buyer-owned commercial license, with exclusivity available.