Short answer
Room-setup data shows a person moving chairs, tables and stage elements into a layout, often with a partner and a floor plan. The useful signal is whole-body: stance, lift, carry path and coordination, so video should be paired with body pose or inertial data. Commissioned recording at event venues is the usual route; video alone gives no robot actions or load forces.
The buyer's problem
Setting up a banquet hall, classroom or conference room is physical work with simple goals and complicated execution. A floor plan says forty round tables of ten; the crew has to work out the order, where to stack, which aisles to keep clear, and how two people carry a folding table through a door. The objects are large, awkward, sometimes heavy, and most of them are rigid but articulated: folding legs, stacking chairs with interlocking frames, wheeled risers.
For a humanoid, this is a whole-body, long-horizon, spatial-reasoning task. Much of its difficulty is not in grasping but in deciding where to put things and how to move large items through space. Buyers should treat it differently from tabletop work.
Task anatomy
- Read the layout. Take a floor plan, a photo or a verbal order; identify counts and positions.
- Stage materials. Locate stacks, trolleys and storage; decide transport order.
- Move items. Carry chairs (stacked or single), roll tables on edge, lift one end, coordinate with a partner.
- Place. Align to a grid, rows or a head table; check spacing and aisle widths.
- Adjust. Open folding legs, lock them, level a table, straighten chairs.
- Finish. Add linens, signs, cables; walk the room to verify; undo the process for teardown.
Failure modes include blocked exits, uneven spacing, tables placed in the wrong row, items dragged across the floor, a collision with a door frame and an unstable stack. Compare the carry-specific side in manual load handling and carrying.
What published work suggests about whole-body capture
Humanoid learning from human whole-body motion is an active area. HumanPlus built a pipeline in which a low-level humanoid policy was trained in simulation on 40 hours of human motion data, then shadowed a person's body and hands in real time from a single RGB camera, and learned skill policies by behavior cloning with egocentric vision, with up to 40 demonstrations per task and 60-100% success on its tasks, on a customized 33-DoF, 180 cm humanoid (Fu et al., 2024). Its tasks are not furniture handling, and the figures are what that study reported on its platform.
Body-model fitting is a route to pose from video. SMPL-X is a body model with articulated hands and an expressive face, learned from thousands of 3D scans, and SMPLify-X fits it to a single RGB image (Pavlakos et al., 2019). Single-view fitting struggles when a table hides the person or the person leans over a large object, so room setup is a hard test for it. See full-body pose from workplace video.
For large-object tasks, the limit is often not pose but contact: where hands grip a tabletop edge, how force passes through the arms, and whether a partner carries half the weight.
Data needs by sub-task
| Sub-task | Helpful modality | Common gap |
|---|---|---|
| Layout understanding | Top-down or wide room view, plan document | Plan not linked to the video |
| Carry paths | Multi-view video, floor map with coordinates | Camera does not cover doorways and corners |
| Two-person lifts | Both bodies in frame, partner roles labeled | Occlusion between carriers |
| Placement alignment | Fixed overhead, per-item final position | Alignment not measured |
| Chair stacking and unstacking | Close view of contact points | Hidden by the carrier's body |
| Table unfolding | Hands and mechanism visible | Lever and lock details out of view |
Use head-mounted views when the robot is expected to act from a person-like perspective, and fixed wide views for layout. Mixed views follow the logic in egocentric vs exocentric workplace video.
Sourcing
Two-person lifts overlap with handover and collaborative tasks, and a full setup is a long-horizon task that also needs mobile manipulation demonstrations.
Event venues, hotel banquet teams, convention centers, schools and facility-services contractors do this work daily. Existing footage is rare. Security cameras in function rooms may exist, but they show a ceiling view of guests as well as staff, and were not set up for training. A commissioned session is the realistic path: the venue provides the room, furniture and crew between events; you provide the capture plan.
The key design choices are the room and furniture set (round tables, rectangular tables, stackable chairs, staging), the layouts to build, whether crews work in pairs, and which body-capture method is acceptable to employees. Wearable sensors need more consent and comfort planning than cameras. Detail in facility services contractors as suppliers.
Because layouts and venue types vary, describe the setup tasks and venue types to SourceX and it will look for suitable US businesses and discuss whether commissioned recording is feasible.
Illustrative package
# Illustrative only - synthetic values, not an available dataset
package: room_setup_pilot
venue: event_venue_009, ballroom 20 m x 12 m, between bookings
layouts: [banquet_rounds_x8, classroom_rows_x6, theatre_chairs_x40]
crew: 6 staff (setup_H01..H06), working solo and in pairs, written consent for model training
furniture: round tables, folding rectangular tables, stackable chairs, 2 carts
views: overhead_wide x2; doorway_view; head_mounted on 2 crew; optional wearable IMUs
labels:
- phase: read_plan / stage / move / place / adjust / verify
- object_id per table and chair stack; final_position_cm vs plan
- carry_mode: solo / paired / cart
- events: collision, blocked_aisle, re-placement
excluded: lift forces, grip pressure, robot actions
rights: training and evaluation; no redistribution
What this data cannot support
- Robot actions and loads. Video contains no joint commands, and carry loads and grip forces are not visible unless measured. Humanoid payload limits differ from a person's.
- Body-to-body mapping. A human crew bends, twists and uses momentum in ways a humanoid with different joints and balance may not. See the human-to-humanoid embodiment gap.
- Safety near people. An empty ballroom omits guests, children and moving obstacles.
- Pose accuracy under occlusion. Estimated body pose can be wrong when the person is hidden behind a table. Ask for confidence flags.
- Layout intent. The reason a planner chose a layout is in documents and conversation, not in clips.
Questions to ask suppliers
- How many rooms, venue types and furniture models are covered?
- Is the floor plan or layout order captured and linked to each clip?
- Is final placement measured, or only visible in video?
- Are paired carries recorded with both people visible?
- Which body-capture method was used, and how are occluded frames flagged?
- Were guests, event staff from other companies or attendees ever in frame?
- What did crew members consent to, including use of body shape and movement?
- Are collisions, blocked aisles and re-placements kept?
- How are cameras synchronized and calibrated across a large room?
- Can a separate venue or layout be held out for evaluation?
Discuss your requirements
Discuss your humanoid training-data requirements with SourceX by describing the room types, tasks, capture methods and permitted uses. SourceX reviews the requirement and follows up about potential sources; 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, and the catalog page on robotics and embodied AI.
Describe the workplace demonstration data your robotics project needs
Sources
- HumanPlus: Humanoid Shadowing and Imitation from Humans — humanoid shadowing from human motion data and RGB input.
- Expressive Body Capture: 3D Hands, Face, and Body from a Single Image — body model with hands and fitting from a single image.