Short answer
Installing gaskets, seals and hoses means seating flexible parts into rigid assemblies, where success depends on feel, alignment and force. Video can show sequence, orientation and visible seating. It cannot record the push, stretch or resistance, so control training needs robot or teleoperation data with force sensing, not footage, work instructions or recognition labels alone.
The buyer's problem
A rigid peg in a hole has a pose. A rubber seal in a groove has a pose plus a state: stretched in places, compressed in others, partly seated, twisted or fully home. Hoses add length, sag and the way the free end flops while the other end is pushed on. People handle all of this with their fingers, adjusting by feel, and most of what they sense never reaches a camera.
Installation tasks like these appear in automotive and equipment assembly, and a buyer may reasonably assume that video of experienced workers is the right starting point. It is, for some purposes. It is not enough for others, and the gap is larger than for rigid-part assembly. This page explains where that gap sits and how to plan around it.
Task anatomy
Variants differ, but the common patterns are:
- Seating a gasket or O-ring: locating the part in a groove or on a flange, pressing it progressively around the perimeter, checking that it lies flat and is not twisted.
- Pushing on a hose: aligning the end to a barb or fitting, pushing against resistance, sometimes using lubricant, then positioning a clamp or clip.
- Routing and clipping: laying a hose or seal strip along a path and securing it in clips or channels. This overlaps with wire harness assembly.
- Inserting profile seals: pressing a long strip into a channel, working from one end, trimming or joining as needed.
- Verifying: looking along the seal, tugging a hose, or running a leak or fit check where the process includes one.
Failure modes include a twisted gasket, a seal pinched at a corner, a hose not fully seated on its barb, a clamp placed over the wrong zone, a kinked hose and a seal that appears seated but lifts later. Several of these are invisible in a single frame and show only as a small shape change along a sequence. Parts found wrong after installation go to repair, which is covered in rework and repair operations.
What the data can support
Video of experienced installers can support understanding of order, hand placement and the visual cues people use to decide that a part is home. For perception, a dataset can label part state: unseated, partially seated, seated, twisted. Labeling deformable parts is its own challenge. ZeroWaste targets highly deformable, often translucent objects in extreme clutter on a recycling sorting line (Bashkirova et al., 2021). That is waste sorting, not installation, so it is relevant only to how deformable-object segmentation can be annotated.
For the manipulation side, RH20T is useful as an illustration of what a contact-rich dataset records. It contains more than 110,000 contact-rich real-world robot manipulation sequences, with visual, force, audio and action data for each sequence, and each robot sequence is paired with a human demonstration video and language description (Fang et al., 2023). It does not cover seal or hose installation. The relevant lesson is the combination: images, force and actions together, with the human video as a companion.
Why force and compliance matter
A seal seats when the operator presses it past a resistance point and feels it drop in. A hose seats when it stops moving on the barb. Those events are defined by force and displacement, not appearance. Two operators may push with very different forces and arrive at the same visible outcome; a third may push too little and look identical until the part fails later. Without force or tactile data, a model cannot distinguish these cases. See force-torque data for what the signal is and snap-fit and press-fit assembly for the closest rigid-part analogue.
What data helps each part
| Part of the task | Data that helps | Typical source |
|---|---|---|
| Order and routing | Overhead and oblique video with step labels | Existing footage or commissioned recording |
| Part state (seated or not) | Close views with state labels | Commissioned recording and annotation |
| Insertion force | Force sensing on tool, part or rig | Instrumented commissioned capture |
| Final check | Leak, pull or fit-test records | Quality records |
| Part variants | Part numbers, materials and hardness | Engineering data |
| Hand technique | 3D hand tracking from a calibrated rig | Commissioned recording |
Illustrative package
# Illustrative only - synthetic values, not an available dataset
package: deformable_install_v0
site: site_048 (synthetic)
tasks: [seat_gasket_in_groove, push_hose_on_barb, clip_profile_seal]
parts: [gasket_G1, hose_H4, profile_seal_P2] # synthetic part ids
streams:
overhead_video: 1080p, 30 fps
close_view: seating region in frame, 60 fps
force: workpiece-side sensor or instrumented tool, own clock, optional
labels:
steps: place | press | seat | route | clip | verify
part_state: unseated | partial | seated | twisted | pinched
verification: pass | rework, source named
metadata: material_family, lubricant_used yes/no, operator_id operator_H02 (pseudonymous)
sync: shared clock or sync event with measured drift
excluded: robot actions, joint states
rights: training and evaluation; part designs stated separately
Limits: what this data cannot support
Video of installation does not capture the force an operator applies, the resistance a part returns or the subtle changes a hand feels. No amount of footage, instructions or quality records replaces robot, teleoperation or instrumented-handling data for control, and recognition or pose labels do not either. Human fingers also differ from robot hands in compliance and sensing, which limits transfer; see the human-to-humanoid embodiment gap.
Task-specific limits to keep in view:
- Material variation. Rubber hardness, temperature and lubrication change how parts behave.
- Latent defects. A seal that looks seated can fail a leak test; outcome labels need the test result.
- Part geometry. One groove profile does not represent others.
- Small samples. Rare errors, such as a pinched seal, may appear too few times to learn from.
Permissions and provenance
Assembly sequences, part designs and supplier names are often confidential, and automotive and equipment programs commonly carry customer terms. Workers are identifiable by face, hands and badges. Redaction or removing identifiers does not settle permissions; blurring a part number does not clear rights in the design visible around it. Ask who can license the footage and what employees were told; the handling problems of hose-end fittings overlap with connector mating and precision insertion. This is not legal advice. See automotive final assembly for line-level context and manufacturing data confidentiality.
Questions to ask suppliers
- Which gasket, seal and hose types, materials and fittings are covered?
- Do recordings include a close view of the seating region?
- Is any force or pressure signal recorded, and where is the sensor?
- How is a seated part defined, and who judged each installation?
- Are leak, pull or fit-test results linked to each unit?
- Are twisted, pinched and partially seated cases present, or only good installs?
- Did lubricants or tools vary, and are they recorded?
- Which part designs or supplier references must be excluded?
Discuss your requirements
If flexible-part installation is on your task list, describe the parts, sensing and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources among US businesses, checks the data and the supplier's licensing permissions, and coordinates license and delivery. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match. See the humanoid training data guide and the manufacturing hub; related fabric handling is covered in sewing and textile operations.
Describe the workplace demonstration data your robotics project needs
Sources
- ZeroWaste Dataset: Towards Deformable Object Segmentation in Cluttered Scenes — deformable-object segmentation; labeling reference only.
- RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot — contact-rich data with force and action streams, as an illustration of recorded modalities.