Restaurant Service: Bussing, Serving and Table Reset Demonstrations

The buyer's problem

Clearing and resetting a table looks like an easy task and behaves like a hard one. The scene changes every time: half-finished plates, stacked cups, a napkin hiding a fork, a spilled glass, a chair pulled out at a random angle. The worker also moves through a room with people, trays and doors, so the task is mobile manipulation with a long tail of small decisions, not a fixed-station pick-and-place.

A team building a service-capable humanoid needs examples of the full cycle and its variations. Restaurant groups know the workflow well, but their usual recordings (security cameras, training clips) do not show it in a form a model can learn from. This page explains what to ask for and how to collect it without recording guests.

Task anatomy

A bussing and reset cycle has these parts, and each deserves its own label:

1. Approach and assessment. Walk to the table, read its state (occupied, finished, still eating), decide what to take.
2. Clearing. Pick tableware in a sensible order: glasses, cutlery, plates. Stack plates and move scraps. Decide how to hold more than two items at once.
3. Carrying. Balance a tray or stacked load while moving between tables, around chairs and through doorways.
4. Surface cleaning. Wipe the table, check for crumbs, sanitize if required by the venue's rules. See wiping and surface cleaning.
5. Reset. Replace cutlery, napkins, condiments and menus in the venue's layout, straighten chairs.
6. Hand-off. Deliver items to the dish pit or cart, where dishwashing and dish racking takes over.

Serving runs the sequence in reverse: collect from a pass, carry, place, adjust. The failure modes differ: a dropped plate while bussing, a spill while serving, a wrong table, an unreachable far corner.

What published work shows about related data

The public egocentric dataset most often used as a labeling reference here records kitchens, not dining rooms. EPIC-KITCHENS-100 contains 100 hours of head-mounted recordings across 45 kitchens, with roughly 90,000 annotated actions; its first annotation stage was "pause-and-talk" narration by the person who filmed, followed by transcription, parsing and temporal segmentation (Damen et al., 2020). It is cited only for its capture and labeling approach; it contains no dining-room or front-of-house footage. That narration-first workflow is worth adapting for front-of-house tasks, because the person who just did the task knows why they moved a glass before a plate.

Mobile ALOHA is a capture reference for mobile manipulation: it uses whole-body teleoperation, in which the operator is tethered and backdrives a mobile base, with 50 demonstrations per task, and reports that co-training with static ALOHA datasets raised success by up to 90% (Fu et al., 2024). That is a robot-teleoperation result, not a human-video one, and it is not evidence about restaurant or dining-room tasks. It is cited for how its demonstrations were captured: on the robot, with other data used to help, which is the role restaurant human video would complement rather than replace. For the trade-off, see human demonstrations vs teleoperated robot demonstrations.

Neither study is a restaurant-service dataset, and neither covers dining rooms or table service.

Why staged recording is the realistic route

Live service recordings capture guests' faces, conversations, children, payment cards on the table and sometimes medical or religious context from what people say. Workers in view have not necessarily consented to their data being used for AI. Venues may have signage for security cameras, but that notice describes security use, not model training. This is not legal advice; ask your own counsel and the supplier what notice and consent applied, and see worker consent and identifiable people.

Staged sessions avoid the guest problem. A restaurant closes a section, or a venue runs sessions before opening, and employees who have agreed to be recorded perform scripted and unscripted resets with real tableware and realistic mess. The cost is realism: the pace is slower, nobody is waiting, and customers do not interrupt. Specify some interruptions deliberately, such as a staff member walking through or a chair blocking the route. For the general trade-off, read staged vs live operation recording.

Existing footage has a narrow role here. Restaurant groups sometimes have staff training videos showing the reset sequence in an idealized form. Those clips can help with step language and sequence, and rarely show recovery behavior.

Illustrative capture package

# Illustrative only - synthetic values, not an available dataset
package: dining_reset_pilot
venue: restaurant_site_004, section closed to guests, off-hours
scenes: 12 table setups x 3 states (just finished, partially eaten, spill present)
demonstrators: 5 employees, ids server_B01..B05, written consent for model training
cameras:
  head_mounted: 1 per demonstrator, 30 fps
  fixed_exocentric: 3, covering table, aisle and pass
  mobile_follow: optional chest-mounted wide view
labels:
  - step: approach / clear / carry / wipe / reset / hand_off
  - object_set: tableware categories present, count per table
  - carry_mode: two-hand stack, tray, one-hand
  - outcome: complete / dropped_item / spill / retried
  - language: one narrated sentence per step, recorded during the task
excluded: robot actions, forces, tray weight (unless logged separately)
rights: training and evaluation; no redistribution; delete on request

Quality criteria a buyer can set

• Mess diversity. Count distinct tableware shapes and materials (glass, ceramic, metal, plastic), not only clip counts.
• Failures included. Drops, near-drops, spills and retries must remain in the dataset with a label. Deleting them leaves the dataset with no examples of recovery.
• Hands and tray visible. Sample clips and check the proportion in which both hands and the carried load are in frame.
• Consistent clock. If multiple cameras are used, check that frame offsets are documented.
• Held-out set. Reserve a venue or table configuration for evaluation, not a random slice of clips. See fragile item handling for how breakable objects affect labels.

What this data cannot support

• Robot actions. Human video has no joint states, commands or gripper signals. Bussing needs a retargeting or co-training plan.
• Forces and weights. Plate stack weight, grip pressure on a wet glass and the tray's balance are invisible. Log weights if you need them.
• Crowd behavior. A staged session with no guests says nothing about how a robot should negotiate people in a busy dining room. That is a navigation and safety question outside this data.
• Cleanliness judgments. Whether a table is "clean enough" depends on the venue's standard. Capture the standard as a written rubric and label outcomes against it.
• Generalization across venues. One restaurant's layout, tableware and workflow cover one restaurant.

Questions to ask suppliers

1. Were guests present at any point during recording, and how were they excluded from frame?
2. What did employees consent to, and does it cover model training and the retention you need?
3. Is the tableware real and varied, or a small repeated set?
4. Are drops, spills and retried attempts kept and labeled?
5. How were steps segmented, and who wrote the language descriptions?
6. Do clips include walking between tables with a load, or only the table itself?
7. What camera mounts were used, and can head-mounted and fixed views be aligned in time?
8. Does the venue's cleaning and sanitation standard appear as a written rubric?
9. Which items or locations were blurred or masked, and what did that remove?
10. Can a separate venue be reserved for evaluation?

Discuss your requirements

If you need front-of-house service demonstrations, describe the tasks, camera views and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources, including whether a commissioned collection is feasible; 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 page for video recordings. Related collection is described under first-person video of skilled manual work. If your task also involves moving through a space, see mobile manipulation demonstrations.

Describe the workplace demonstration data your robotics project needs

Sources

• Rescaling Egocentric Vision — kitchen video scale and narration-first annotation workflow.
• Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation — mobile bimanual teleoperation demonstrations and co-training.