Short answer
Bicycle and light-vehicle repair is an accessible domain for demonstration video: tasks are short, parts are small, and public research has already recorded bike repair with multiple synchronized cameras. Shops and micromobility fleets may be able to supply or host recordings. The video supports step understanding and perception, but not robot actions or forces, and e-bike electronics add safety and diagnostic aspects video alone does not cover.
Why this domain keeps showing up
Bicycles are a convenient testbed. The workpiece is small enough to put on a stand under controlled lighting, the hardware is standardized enough to recognize across brands, and the tool set is compact: hex keys, wrenches, tire levers, chain tools, pumps. A recording can be made in a shop without industrial access rules or plant shutdowns.
Public precedent exists. Ego-Exo4D recorded skilled human activities, including bike repair, simultaneously from an egocentric camera and multiple exocentric cameras, and added multichannel audio, eye gaze, 3D point clouds, camera poses, IMU data and expert commentary, with benchmarks that include fine-grained activity understanding and proficiency estimation (Grauman et al., 2023). That makes it a reference for what a well-instrumented repair recording can contain. It is not a dataset of workplace shop recordings made for robot control, and its bike-repair portion should be assessed separately from the overall figures the paper reports.
If your question is the wider landscape of repair video, the repair video demonstrations page comes first. For vehicles beyond bicycles, see automotive service tasks.
Task anatomy
A plausible task list for supplier conversations. Ask technicians to confirm and extend it.
- Wheel and tire work. Removing a wheel, levering a tire, patching or replacing a tube, reseating, inflating.
- Drivetrain. Chain removal and installation, cassette or chainring work, derailleur adjustment.
- Brakes. Pad replacement, cable or hose work, alignment, adjustment.
- Bearings and headset. Disassembly, cleaning, regreasing and reassembly in a specific order.
- Fit and fastening. Tightening stems, seat posts and bolts to specified values.
- E-bike additions. Battery handling, connector inspection, display diagnostics. These add electrical hazards and manufacturer-specific procedures.
Failure modes include stripped threads, seized fasteners, wrong tool size, dirty chains, cross-threaded parts and misadjusted brakes. Some are visible; others only show up as a tool slipping, a creak or a part that fails a ride test.
What different footage can support
| Footage | Supports | Does not support |
|---|---|---|
| Tutorial or marketing video | Step names, tool vocabulary | Reliable hand-object detail; rights are usually restricted |
| Shop security video | Context, task timing | Usually the wrong viewpoint; faces and customers |
| Technician body-worn recordings | First-person view of the work | Stable framing; hands often occluded |
| Purpose-built multi-view recording | Step segmentation, hand-object perception, proficiency comparisons | Robot actions, forces |
| Instrumented tool or rig recordings | Force and motion signals | Breadth across shops |
Consult technician body camera footage for the first-person case, and hand tool use demonstrations for the broader tool-use setting.
Illustrative recording package
This example is synthetic. Counts, views and fields are buyer decisions.
# Illustrative only
task_families: [tire_and_tube, brake_pad_swap, chain_service]
bikes: types_buyer_defined # e.g. commuter, e-bike
sessions:
- session_id: s_0007
shop: shop_004
mechanic: mech_A09 # pseudonymous
views: [head_ego, fixed_front, fixed_side]
clocks: shared_marker_at_start
outcome: completed | needed_redo | abandoned
labels:
- step_segments with tool names
- part names as the mechanic says them
- hands_visible flag
- mistake_or_redo markers
- spoken_narration_transcript (optional)
excluded: [robot_actions, torque_values_applied, hand_force]
Where torque matters, ask whether the shop used a torque tool and whether readings were recorded. Without them, the video cannot tell a model how tight is tight. See instrumented tools and torque specs and fastening procedures.
Where the video helps most, and where it misleads
Bike work suits visual tasks. Identifying a part, recognizing which fastener comes next, and noticing a misrouted cable are all things a camera records well, and mechanics can narrate their reasons as they go. Label scheme design can borrow from tool affordances and grasp annotations, since much of the work is about where a tool engages a part and how the hand holds it.
Two traps are worth naming. First, a clean stand-mounted recording of a wheel change looks tidier than a roadside repair or a crowded shop floor, and a model tuned only on the tidy version may degrade quickly in the other. Second, proficiency differs sharply: a shop veteran and an apprentice can complete the same job with very different efficiency. Ask for experience level per mechanic so you can stratify, instead of treating all recordings as equal.
Sourcing routes
Independent bike shops, shop chains, bike-share and scooter fleets and cargo-bike operators perform repetitive repairs at volume. Fleet operators in particular tend to have maintenance workflows, parts catalogs and work orders that can link to video if recording is added. Existing footage is unlikely to be suitable, so commissioned recording at cooperating sites is the main route; compare in existing footage versus commissioned recording.
Recordings of mechanics at work, and sometimes of customers or their property, require consent and notice that the site designs. See worker consent and identifiable people. This is not legal advice.
Limits: what this data cannot support
- Human hands, not robot hands. Bike repair relies on dexterity and strength that a humanoid hand may not match. See the embodiment gap page.
- No forces. Video does not show how hard a nut was tightened, a spoke tensioned or a tire pushed onto the rim.
- Easy-domain bias. Because the task is accessible, results from it may overstate how well a method transfers to harder industrial repairs.
- Brand and model variety. Standards differ across components; one shop's inventory is a narrow sample.
- E-bike safety. Electrical and battery procedures are manufacturer-specific. Video of someone doing a job is not a statement of the right procedure.
- Control. Training a humanoid to do these tasks needs robot or teleoperation data. Human video may help perception, step understanding and planning.
Questions to ask suppliers
- Which repair families and bike types are covered, including e-bikes?
- Are recordings of full jobs, or only highlight segments?
- How were mechanics briefed, and did they narrate?
- What camera setup was used, and are clocks synchronized?
- Are mistakes, redo steps and tool swaps retained?
- Are torque-tool readings or part numbers recorded alongside video?
- Were customers or their bikes visible, and what notice was given?
- Can records of the job (work orders, parts) be linked to each video?
- Who can authorize licensing: the shop owner, the franchisor or the fleet operator?
Discuss your requirements
If bike and light-vehicle repair belongs in your demonstration plan, describe the tasks, camera setup and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material or could host new recordings, and follows up about potential sources; commissioned collection is discussed case by case. A request does not guarantee a match. Begin at the humanoid training data guide for the broader map or the maintenance cluster for neighboring pages. The catalog entry on egocentric video of manual work is related.
Describe the workplace demonstration data your robotics project needs
Sources
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — public multi-view skilled-activity recordings including bike repair, with gaze, IMU and proficiency benchmarks.