Short answer
Cart and trolley demonstrations show how people steer wheeled loads through corridors, doors, elevators and crowds. Video captures route, turning, stopping and negotiation with people, but not the push force, load mass or floor friction that determine how the cart behaves. Buyers typically need load and surface records, whole-body or mobile-base robot data, and varied sites.
Task anatomy
Pushing a cart looks simple until you list what a person does. In a hotel, a housekeeper moves a loaded linen cart down a corridor, parks it outside a door, props the door, and edges the cart through a doorway. In a hospital, a porter pushes a supply trolley past waiting patients, into an elevator and around a tight corner. In a warehouse, a picker pulls a multi-tier cart along an aisle, stops at locations and reorients it. In retail, someone shuttles a stock roll cage from a delivery bay to the sales floor.
Sub-skills recur across these settings.
- Engage. Choose a grip point, set the casters, release a brake.
- Accelerate. Overcome static friction, adjust to load mass and floor surface.
- Steer. Turn around corners, correct drift from a bad caster, handle a slope or threshold.
- Negotiate. Yield to people, signal, slow near blind corners.
- Dock and park. Align at a station, set brakes, leave clearance.
- Transition. Pass through doors, onto elevators or over thresholds.
Failure modes include a caster jammed on a seam, a cart overloaded to one side, collisions with door frames, a blocked corridor and rolling away on a slope.
What video shows and what it misses
| Aspect | Visible in video | Missing |
|---|---|---|
| Route and turning geometry | Yes | Exact path unless localized |
| Speed and stops | Yes | Push force over time |
| Interaction with people | Yes | Intent unless labeled |
| Door and threshold handling | Yes | Door resistance, threshold impact force |
| Load condition | Partly | Mass, center of mass |
| Floor type | Partly | Friction, slope |
| Hand placement | If camera sees hands | Torque at handles |
The combination of force and navigation is why cart pushing is a separate topic from carrying or walking alone. The robot must regulate force at the handles while it plans a path among people, and a mismatch between load mass and push effort shows up as lurching or stalling. Video can inform path choice and social behavior, but the force side comes from other data, as covered in when video is not enough.
What published work offers
Mobile ALOHA describes whole-body teleoperation of a bimanual mobile manipulator: the operator is tethered to the robot and backdrives its mobile base. It reports 50 demonstrations per task, and co-training with static ALOHA datasets raised success by up to 90% (Fu et al., 2024). It is relevant because it shows a way to collect base and arm motion jointly, rather than because it studied carts. The result applies to its own tasks and hardware, not to any cart-pushing program, and it is a data point for a mixed strategy of small task-specific datasets plus co-training. See human demonstrations vs teleoperated robot demonstrations.
BEHAVIOR-1K defines 1,000 everyday activities across 50 scenes in the OmniGibson simulator, with activities drawn from time-use surveys and preference ratings (Li et al., 2024). It can supply task vocabulary and simulated scenes for household-scale mobile tasks, but it is not a source of real cart-handling data.
Which data helps each part
For route and social behavior. Head-mounted and fixed video showing people yielding, signaling and sharing corridors; see mobile manipulation demonstrations.
For doors and elevators. Close views of the cart-door interaction; see doors, drawers and cabinets.
For load and force. Instrumented handles, load cells or motor-current logs from powered carts; a record of load mass and floor type per run.
For setting-specific routines. Hotel housekeeping room turnover and parcel delivery handling show where carts sit in a larger workflow.
For robot-side data. Only robot or teleoperated data supply a humanoid's whole-body commands, balance and base interaction. Human footage does not.
Supplier types
Hospitals, hotels, warehouses, retail stores and facility-services contractors all move carts every day. Hospitals add patient privacy questions, so footage that shows patients or health information may be unusable; hotels add guest privacy. This is not legal advice. Warehouses are easier to record but may restrict cameras near vehicles. Existing footage from security or training systems rarely holds useful hand views. See existing footage vs commissioned recording.
Capture considerations
- Viewpoints. Head-mounted for hands and handle contact; a chest or hip-height forward camera for path; a fixed view at tight points such as doors.
- Localization. Add a map or fiducials so path can be reconstructed.
- Load record. Log cart type, load mass and distribution, wheel type, floor surface and slope.
- Handle sensing. An instrumented handle or wrist sensor, where allowed.
- People in view. Plan consent for staff, patients, guests and visitors; background individuals cannot all be asked for permission, so choose times and routes that limit exposure.
- Safety. Sites decide where recording is allowed, especially in clinical corridors.
Illustrative specification
# Illustrative only - synthetic values
task_family: cart_pushing_v0
carts: [linen_cart_A, supply_trolley_B, roll_cage_C]
cart_log: [wheel_type, load_mass_class, load_distribution, brake_type]
routes: [corridor, doorway_with_threshold, elevator, tight_corner, slope_5pct]
views: [head_mounted, forward_hip_camera, fixed_at_door]
sensors: [optional handle_load_cell, optional imu_on_cart]
labels:
phase: [engage, accelerate, steer, yield, dock, transition]
event: [stop_for_person, door_hold, collision, recovery]
outcome: [clean, retry, blocked]
metadata: [site_id, operator_id (pseudonymous), floor_surface, traffic_level]
rights: "staff and visitors in view addressed in site agreement"
Cart pushing versus carrying
Carrying puts the load on the body; pushing puts it on wheels. The robot that carries needs payload and balance data, covered in manual load handling and carrying. The robot that pushes needs handle force, base motion and path data. Some jobs mix them: a worker lifts a bag onto a cart, pushes the cart, then lifts it off. Scope each phase separately so the dataset does not blur two different control problems.
Limits and extra preparation
Human cart pushing involves whole-body motion and weight shifting that a humanoid will not copy directly; see the embodiment gap. Datasets from one hotel or hospital may reflect that site's carts and layout. Rare events, such as a blocked doorway or an emergency, are scarce. Evaluation sets should be held out by site and cart type. Pace may differ from what a new robot can safely do.
Questions to ask suppliers
- Which cart types, loads and floor surfaces appear, and are they logged?
- Are routes with doors, elevators, thresholds and slopes included?
- How often do people appear in the footage, and how are they covered by consent?
- Do any clips include handle force, load cell or motor-current data?
- Which camera views show the hands on the handle?
- Are retries, collisions and blocked-path recoveries included or edited out?
- How many sites, shifts and operators are represented?
- Which areas could not be recorded, such as patient rooms or secured zones?
- Is any footage from existing security or training systems, and if so how was it recorded and for what purpose?
Discuss your requirements
If your plan includes moving wheeled equipment, describe the settings, loads and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources, and can scope a commissioned collection of new recordings if no existing asset fits; a request does not guarantee a match. For wider navigation, see the workplace task demonstration hub and the humanoid training data guide. The catalog entry on video recordings is related.
Describe the workplace demonstration data your robotics project needs
Sources
- Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation — whole-body teleoperation of a mobile manipulator and co-training results.
- BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation — simulated everyday activities and task vocabulary.