Short answer
Video of tool use in tight spaces shows how skilled people reach, brace and sequence work when access is poor, but the same constraints that define the task hide it from cameras. Wearable and multi-view capture help; both leave out joint commands and forces. A humanoid's own body and reach differ, so this data supports perception and planning research rather than control training on its own.
Why access is the real task
In many repairs the hard part is not turning the screw. It is getting a hand, a tool and a line of sight to the screw at the same time: behind a panel, under a dashboard, inside a cabinet, around a pipe. Technicians solve this with posture, tool choice, extensions, mirrors and sequence changes that never appear in a procedure.
For a humanoid researcher this is also where the platform is most different from a person. Link lengths, wrist range, hand size and head-camera position all determine whether a human strategy is even feasible. Data from people shows which strategies exist, and the human-to-humanoid embodiment gap determines how much of that carries over. Treat this page as a guide to buying evidence about constrained work, not as a promise that it transfers.
What the task looks like
Constrained tool use has features that a standard bench demonstration lacks. The list below is a framing for conversations with a supplier, not a fixed taxonomy.
- Blind or partly blind operation. The tool tip is out of sight, and the person works by feel, from a mirror or by moving their head.
- Non-standard grips and postures. Choking up on a handle, using the wrist at the edge of its range, supporting a tool with the free hand or forearm.
- Tool substitution. Swapping to a shorter driver, ratchet, offset wrench or flex extension because the default does not fit.
- Staging and retrieval. Dropped fasteners, parts held in teeth or on a tray, lighting added mid-task.
- Repeated attempts. Positioning, failing, readjusting. Much of the real duration lives here.
The hand tool use demonstrations page covers open-bench tool use, and the disassembly and reassembly sequences page covers order of operations. This page concentrates on what happens when access constraints control the camera placement itself.
Capture options and what each shows
| Capture approach | What it adds | What it loses |
|---|---|---|
| Head-mounted camera | The person's view of the work, including mirror use | Often blocked by the body in tight spaces; heavy motion |
| Chest or shoulder camera | Stable view of arms and workspace | Hands may be hidden by the workpiece |
| Wrist or tool-mounted camera | Close view of tool tip and fastener | Unusual viewpoint; tool vibration |
| Fixed external camera | Whole-body posture and context | Occluded by the structure being worked on |
| Multiple synchronized views | Cross-checking of hand position | Setup time; fixed cameras cannot enter small cavities |
| Body-worn IMUs or gaze tracker | Posture and attention signals | Sensor drift; extra preparation |
Ego-Exo4D is a useful reference for combining views: it records skilled activities, including bike repair, with simultaneous egocentric and multiple exocentric cameras, and adds multichannel audio, eye gaze, 3D point clouds, camera poses and IMU data (Grauman et al., 2023). It is a capture and labeling reference only. Bike repair on a stand is not a confined-space setting, so do not read its results as evidence about cramped-access work.
Illustrative recording specification
This is synthetic, for discussion. Every parameter is a buyer decision.
# Illustrative only
task: replace_component_with_restricted_access
subject_of_study: access_strategy_and_tool_choice
episodes: buyer_defined
cameras:
- id: cam_head # egocentric, with timestamp
- id: cam_wrist # optional
- id: cam_ext_1 # fixed, wide, calibrated if possible
- id: cam_ext_2 # second angle, may be partly occluded
sync: shared_clock_or_visual_sync_marker
annotations:
- segment: [approach, position, engage, actuate, withdraw]
- posture_label: free_text_plus_category
- tool_swap_events: true
- visibility_flag: tool_tip_visible | hand_visible | both | neither
- retry_count_per_step: true
excluded: [joint_commands, contact_forces, torque_applied]
If you want to estimate tool motion rather than just recognize steps, plan for tool pose tracking annotation, which needs visibility you may not get in a cavity.
Sourcing
Repair and maintenance providers, appliance and automotive shops, building services firms and field-service organizations do this work every day. Existing footage of it is uncommon, usually partial, and mostly taken for training or dispute documentation rather than for modeling. That makes commissioned recording the likelier route. The technician body camera footage page discusses what body-worn recordings typically contain, and existing footage versus commissioned recording lays out the trade-off.
Because people in tight spaces are often alone, recorded in customer premises or filmed near customers' property, consent and site permission need explicit attention. See worker consent and identifiable people. This is not legal advice.
Limits: what this data cannot support
- No actions or forces. Video shows posture and tool position. It does not carry joint states, commanded motion or how hard the person pushed or twisted.
- Occlusion is built in. The most informative moments, such as fastener engagement, are the ones cameras are least likely to see. Visibility flags help you filter; they do not recover missing data.
- Different reach envelope. A person's strategy may need a wrist angle or elbow position a humanoid arm cannot reach. Filter or label strategies by the kinematic assumptions you hold.
- Motion blur and head sway. Awkward work means abrupt head and body movement, which degrades frames.
- Narrow task coverage. Tight-space jobs vary by vehicle, appliance or building; do not generalize from one assembly.
- Control training. For manipulation policies you still need robot or teleoperation data. Human footage may inform perception, step understanding and planning.
Questions to ask suppliers
- Which jobs and assemblies are included, and how restricted is access in each?
- Which camera mounts were used, and how often is the tool tip visible?
- Were the technicians told to explain their approach aloud, and is that audio kept?
- Are retries, dropped parts and tool swaps preserved, or were they edited out?
- Are cameras synchronized and calibrated, and how is that evidenced?
- How were technician height, handedness and experience level recorded?
- Can recordings be tied to the repair outcome or work order?
- Did the work take place on customer premises, and has the customer agreed to recording?
- What does the supplier do about frames that show faces, plates or customer documents?
Discuss your requirements
If constrained-access tool use matters to your program, describe the tasks, camera setup and permitted uses to SourceX. SourceX looks for US businesses that may hold or be able to record such work and follows up about potential sources; specialist capture partners can be engaged case by case, and a request does not guarantee a match. For the larger picture of supply routes, see the humanoid training data guide and the maintenance cluster. The catalog entry on first-person video of manual work describes a related collection program.
Describe the workplace demonstration data your robotics project needs
Sources
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — simultaneous egocentric and exocentric capture of skilled activities including bike repair; capture reference only.