Short answer
Beverage-preparation demonstrations pair a fixed workstation, machines and liquids with a strict step order, so they suit multi-view recording with step labels and outcome checks. Existing cafe training video can supply sequence and language; commissioned sessions can add viewpoints and failures. Video shows neither pour dynamics, steam or grip force, nor robot actions.
Why drinks are a distinct data problem
A cafe bar is compact, repetitive and highly structured, which makes it attractive for a first manipulation deployment. It is also unforgiving. A drink order is a sequence in which timing matters: shots are pulled while milk is steamed, a lid goes on a cup that has already been filled, a ticket changes mid-sequence. The objects are cups, jugs, portafilters, syrup pumps and buttons that look similar but behave differently, and the liquids are translucent and reflective, which is hard for perception.
A buyer needs data that captures the sequence, the machine interaction and the outcome, and must be clear about the parts a camera cannot see.
Task anatomy
A single espresso-based drink involves these stages, each worth a label boundary:
- Ticket reading. The worker reads or hears an order and selects cup, size and modifiers.
- Grinding and dosing. Fill and level the portafilter, tamp, lock into the group head.
- Extraction. Start the shot, watch timing, place a cup or shot glass beneath.
- Milk handling. Pour milk into a jug, steam it, purge the wand, swirl and tap.
- Assembly. Pour into the cup, add syrups or ice, lid, sleeve.
- Hand-over and clean-down. Place on the counter, wipe the wand, knock out the puck, rinse.
Drip, tea, cold-brew and blended drinks have different sequences, and a dataset for one is not evidence for another. The pouring step in particular is its own technical area; see liquid pouring and dispensing.
Failure modes to capture rather than discard: a puck that does not seat, a cup placed off-center, milk over-steamed, a spill, a mis-pressed button, a ticket that changes while the drink is in progress.
What each recording view contributes
Beverage stations reward multi-view recording because the interesting action is split between hands, machine controls and the liquid surface.
| View | What it shows | What it misses |
|---|---|---|
| Head-mounted (first person) | Hands, cup, tool in hand, rough gaze direction | The machine display if the head turns away; the liquid level from the side |
| Overhead fixed | Layout, tool positions, cup placement | Pour height and jug tilt |
| Side or counter-height fixed | Jug angle, stream, fill level | Hand details behind the machine |
| Close-up on the machine panel | Which button or dial, display state | Anything off the panel |
Ego-Exo4D is a useful reference for this design: it recorded skilled activities simultaneously from an egocentric device and multiple exocentric cameras, with multichannel audio, eye gaze, 3D point clouds, camera poses and IMU data, expert commentary, and benchmarks that include fine-grained activity understanding and proficiency estimation (Grauman et al., 2023). Its activities are not cafe tasks, and it contains no cafe or bar-station footage; it is cited only as a capture reference, because its structure of paired first-person and third-person views plus expert commentary can be adapted to a barista station. Read more in multi-view synchronized task recording.
Kitchen video also gives a labeling model. EPIC-KITCHENS-100 used head-mounted cameras, and its first annotation stage was a "pause-and-talk" narration by the camera wearer, then transcription, parsing and temporal segmentation (Damen et al., 2020). EPIC-KITCHENS is a labeling reference only and contains no cafe or bar-station footage. For a barista, narration is where tacit decisions can surface, such as why the shot was discarded.
Sourcing: existing assets and commissioned sessions
Existing footage. Cafe chains may have onboarding videos, new-hire drink-making modules and recipe cards; ask what exists. They are scripted, clean and one-pass, so they provide sequence, language and a standard to judge against, and they carry little recovery behavior. Check whether the chain or a training vendor owns the content before assuming it can be licensed; third-party production is possible. See corporate training video archives.
Commissioned sessions. A cafe site can be recorded before opening or during a quiet period with employees who have agreed to be recorded. You choose the views, drinks list, step labels and number of repeat orders. Real service recording adds customers at the counter, which raises consent questions that staged sessions reduce, though employee consent still needs checking. This is not legal advice. For that decision, see staged vs live operation recording. Whichever route you use, keep the machine model fixed or recorded, because the controls differ from one machine to the next.
If you want a mix, describe the drink list and machine types to SourceX and it will look for US businesses that hold suitable material and discuss whether a commissioned session is feasible. That is a request for review, not an available stock of recordings.
Illustrative package
# Illustrative only - synthetic values, not an available dataset
package: bar_station_pilot
site: cafe_site_011, machine_model recorded as machine_M3
drinks: [espresso, latte, cold_brew_over_ice, syrup_flavored_latte]
orders: 40 tickets per drink type, incl. 6 with a mid-order modification
demonstrators: 4 baristas (barista_C01..C04), written consent
views: head_mounted 30 fps; overhead 30 fps; side_counter 60 fps; panel_closeup 30 fps
sync: shared time code, offsets in metadata
labels:
step_boundaries: six stages above, plus idle and wait
machine_state: grind / extract / steam / idle, from panel view
outcome: accepted / remake / spilled, judged against a written drink spec
language: narrated sentence per stage, recorded by the barista
scene: cup sizes, milk types, ice presence, lighting condition per session
excluded: robot actions, pressure, temperature, force
rights: training and evaluation; no redistribution
Quality and acceptance pointers
- Sequence integrity. Every drink ticket should map to a complete clip with a start and end time, not fragments.
- Outcome labels against a spec. "Good" must be defined by the venue's drink standard, with a reviewer who is not the person who made the drink.
- Liquid visibility. Test whether the fill level and stream are visible in at least one view for the pouring stage.
- Panel readability. If machine state matters, confirm the display is legible at the recorded resolution and frame rate.
- Held-out configuration. Reserve a machine model, site or barista for evaluation. See success and outcome labels and step segmentation annotation.
What this data cannot support
- Robot control. Recordings of baristas contain no robot joint states or commands. A humanoid or arm must still operate real machines, so robot-side data is needed.
- Contact and fluid dynamics. Tamp pressure, the feel of a portafilter locking in, steam wand resistance and milk viscosity are not visible. Instrumented rigs or robot trials supply them.
- Machine integration. Whether a robot can press a button or read a display depends on machine access, not on video. Some commercial machines also have software controls or safety interlocks that are outside this data.
- Hygiene and food safety. Video can record a wipe-down step, and it cannot certify that a drink station meets health rules.
- Menu and equipment transfer. One machine model and menu does not describe another. Specify the equipment your robot will actually face. For adjacent kitchen work, see commercial kitchen food prep.
Questions to ask suppliers
- Which machine models and grinders appear, and is the model recorded per clip?
- Are the clips complete ticket-to-handover sequences or edited highlights?
- Which views were recorded, and how were they synchronized?
- Are remakes, spills and mid-order changes kept and labeled?
- Who wrote the drink standard used for outcome labels?
- Did baristas narrate during the task, or was narration added later from memory?
- Is the training content produced in-house or by a third-party vendor, and who can license it?
- Were customers, partial faces or payment screens visible, and what was removed?
- How many baristas, sites and cup or milk variants appear?
- Can one site or machine be reserved as an evaluation set?
Discuss your requirements
Describe your drink list, equipment and permitted uses, and SourceX will review the requirement and follow 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 the catalog guide to egocentric video of manual work. You can also compare existing footage with commissioned recording before scoping.
Describe the workplace demonstration data your robotics project needs
Sources
- Rescaling Egocentric Vision — head-mounted kitchen video and narration-first annotation.
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — paired egocentric and exocentric recording of skilled activity.