Short answer
Hospital support data for robots, such as moving linen carts, delivering supplies or resetting an empty room, is realistically sourced as staged demonstrations in areas without patients. Footage from occupied wards raises patient privacy questions that a buyer should treat as diligence items, not assume away. Clinical tasks are out of scope, and video alone gives no robot actions or contact forces.
The buyer's problem
Healthcare facilities involve a lot of physical support work a humanoid could take on: transporting supplies, restocking carts, clearing and resetting rooms, moving linen and waste, delivering meals, wiping surfaces. These tasks are non-clinical, repetitive and physically demanding, which makes them candidates for robotics work.
They also happen in places where patients are. An ordinary ward corridor recording can show a patient on a stretcher, a visitor, a whiteboard with names, a medication cart and a monitor display. The data you want, a worker doing routine logistics, comes bundled with data you cannot responsibly use. That is why this page treats staged collection as the realistic route and not an inconvenience.
Tasks that fit a non-clinical scope
| Task | Setting without patients | Notes for capture |
|---|---|---|
| Linen and supply cart movement | Service corridors, back-of-house lifts | Cart weight, door and lift handling |
| Restocking supply rooms and carts | Central supply, storerooms | Item variety, label reading, shelf heights |
| Discharged-room turnover | Empty room after discharge | Surfaces, bed linen, waste. See janitorial workflow demonstrations |
| Meal tray and kitchen logistics | Kitchen and dock areas | Racks, trays, temperature rules |
| Waste and soiled-item transport | Service routes | Contamination rules, bins and bags |
| Parcel or supply hand-off | Between known drop points | Non-clinical items only; no specimens or patient-linked items |
Exclude anything that touches patient bodies, medication administration, clinical judgment or identifiable records. Humanoid robots in direct patient care are a different safety and regulatory domain that no demonstration dataset sourced through this route addresses.
Why patient privacy shapes the route
In the United States, the HHS summary of the Privacy Rule explains that it protects individually identifiable health information held or transmitted by covered entities, which include health plans, health care clearinghouses and certain health care providers, and it is enforced by the HHS Office for Civil Rights (HHS). A recording of a patient in a care setting can contain such information. Whether a particular recording is covered, and what can be done with it, is a question for the facility's privacy officer and counsel. This is not legal advice.
European guidance on video devices addresses processing of personal data captured by video, including the purpose, notice and data-minimization considerations (EDPB, 2020). Buyers outside the US should ask suppliers how their local frameworks apply to footage exported across borders, even though the suppliers are US businesses.
The practical result is that a facility is unlikely to release ward recordings for AI training, and a buyer who accepts them takes on risk. Blurring faces does not by itself resolve permission questions, because voices, wristbands, room numbers, screens and context can still identify people. A staged session in an empty or closed area reduces that exposure, though staff consent and facility approval still need diligence. See redaction in demonstration video for what redaction removes and costs.
Staged collection design
The general trade-off is covered in staged vs live operation recording.
A hospital can host a session in a decommissioned wing, an empty ward or a service corridor closed to the public, with employees who have agreed to be recorded. Elements to agree up front:
- Area control. Signage, closed doors and a monitor ensuring no patient, visitor or record enters the frame.
- Participants. Environmental-services or logistics staff who have given informed consent specific to model training, and who may withdraw.
- Props. Real carts and linens, with mock supplies in place of controlled or patient-linked items. No real specimen, medication or patient label should be in frame.
- Infection-control rules. Capture what staff would actually do, including gloving and hand hygiene, and record them in metadata.
- Scripted variation. Door held by another person, cart blocked by a trolley, spilled item, elevator wait, low shelf.
A narrower option is a care or senior-living facility's laundry, kitchen or storage area, which has the same logistics without resident rooms. Ask each site what facility rules and approvals apply.
Illustrative package
# Illustrative only - synthetic values, not an available dataset
package: facility_support_pilot
site: hospital_site_006, closed service wing, no patients or visitors
tasks: [linen_cart_to_room, supply_cart_restock, empty_room_turnover, tray_rack_move]
staff: 5 (support_F01..F05), written consent for model training, withdrawal procedure documented
props: mock supplies only; no medication, specimens, patient labels or records
views: head_mounted 30 fps; fixed corridor view; fixed room view; lift_door close view
labels: task_step, load_state, door_state, outcome (complete / blocked / retried)
metadata: glove_use, hand_hygiene_events, cart_load_weight_class, floor_surface
excluded: any clinical task, patient data, robot actions, forces
rights: training and evaluation; no redistribution; deletion on request
What this data cannot support
- Clinical or patient-adjacent work. Nothing here teaches a robot to move or touch a patient, deliver medication or make a clinical decision.
- Robot actions. Human video has no robot joint states or commands. Navigation, door opening and cart pushing still need robot-specific data, see pushing carts and trolleys.
- Real-environment crowd behavior. A closed wing contains no hurrying nurses, distressed visitors or emergency alarms. Behavior around people in a live hospital is not in this data.
- Infection-control proof. A recorded hand-hygiene step is evidence of what was done in a session, not of compliance in service.
- Written-procedure authority. The facility's procedures decide what is allowed; see hospital environmental services procedures for the document side.
Questions to ask suppliers
- Which areas were recorded, and were patients, visitors or records ever within view or hearing?
- Who at the facility, such as privacy officer or counsel, reviewed the recording plan?
- What did staff consent to, and can they withdraw?
- Were any real clinical items, labels or screens present, and how were they excluded?
- How were faces, badges and audio handled, and what did that remove from the training value?
- Are infection-control steps recorded and described?
- What cart and load weights were used?
- Are blocked paths, spills and retries kept?
- Where will the footage be stored, and who may access it during collection?
- Can a separate site or area be held out for evaluation?
Discuss your requirements
Discuss your humanoid training-data requirements with SourceX by describing the non-clinical tasks, setting and permitted uses. SourceX reviews the requirement and follows up about potential sources, and whether a commissioned collection is feasible is discussed case by case; a request does not guarantee a match, and nothing is contracted until a supplier agrees. See also the workplace task demonstration hub, the humanoid training data guide, worker consent and identifiable people, and the use-case page for healthcare admin AI for non-physical healthcare data.
Describe the workplace demonstration data your robotics project needs
Sources
- Summary of the HIPAA Privacy Rule — covered entities and protected health information.
- Guidelines 3/2019 on processing of personal data through video devices (Version 2.0) — personal data in video recordings.