Adhesive Dispensing and Sealing Tasks

The buyer's problem

Dispensing looks like a drawing task: follow a line, lay a bead. The difficulty is that the line is only half of it. The bead's width, height and continuity depend on how fast the nozzle moves, how far it sits from the surface, how much material is pushed out and how the material behaves afterward. A path that is geometrically right can still produce a thin patch, a blob at the start or a gap at a corner.

A buyer asking for adhesive dispensing data is typically trying to learn a continuous-path skill whose outcome appears only on the part. This page covers what manufacturers record around that task, how to design a capture that includes the outcome, and what is missing from the video no matter how well it is shot.

Task anatomy

Hand-held and fixture-based dispensing share a structure:

• Prepare: checking the cartridge or reservoir, purging or priming the nozzle, and cleaning the part surface where required.
• Locate: positioning the part in its fixture and finding the start point, often by eye or a marked guide.
• Dispense: moving the tool along the joint or perimeter while squeezing or triggering, controlling distance and speed.
• Finish: ending the bead, wiping stringing, tooling or smoothing the bead where the process calls for it, and setting the part aside to cure or mating it immediately.
• Inspect: checking bead continuity and placement visually or with a gauge.

Typical failures include gaps, double-ups at the start and stop point, material off the intended line, stringing from the nozzle, air bubbles, and parts mated after the working time has passed. Whether a given adhesive or sealant tolerates a delay is a property of the material and process, and the data should record which one was used rather than assume.

Why the fluid matters

Unlike rigid-part tasks, dispensing changes the object as the task proceeds. The bead is a deformable, time-varying result that depends on material properties such as viscosity and cure behavior. A recording that omits the material name, temperature and cartridge condition leaves out variables that explain differences between otherwise identical-looking demonstrations. Ask what is logged: material identifier, lot, nozzle size, and, where the equipment exposes it, supply pressure.

Hazard handling is also site-specific. Some materials call for ventilation, gloves or skin and eye protection, and a site's rules affect where cameras can stand and what they may be exposed to. Do not assume what a given material requires; ask the supplier for its site rules and consult the material's own safety documentation. See machine safety constraints for recording for the capture-side considerations.

What data helps each part

Many plants already dispense with fixed-automation equipment. The paths in those programs may be useful as geometric references, but they are expressed in a specific industrial robot's coordinates; the caveats are covered in existing industrial robot programs.

How control training enters

Policy-learning methods are trained on observation and action data. Diffusion Policy, for example, represents a visuomotor policy as a conditional denoising diffusion process over actions (Chi et al., 2023). That paper does not cover dispensing, so it is cited only as a method reference. The point for a buyer is the form of the training signal: such a method needs recorded actions in the robot's action space. Footage of a person dispensing, the bead images and quality verdicts supply the observation and outcome sides, and the action side comes from a robot, teleoperation or a handheld capture device.

Illustrative recording package

# Illustrative only - synthetic values, not an available dataset
package: dispensing_demos_v0
process: perimeter_bead_on_housing   # synthetic
material: material_M2 (named by supplier), nozzle_size N-3, lot lot_0144
streams:
  overhead_video: 1080p, 30 fps, part pose marked by fiducials
  nozzle_close_view: 60 fps, lit to avoid glare on the bead
  tool_log: trigger state and pressure if the equipment exposes them, own clock
  tool_pose: optional tracked marker on the handheld tool
labels:
  segments: prime | locate | dispense | finish | inspect
  bead_outcome: continuous | gap | blob | offset | stringing
  gauge_readings: bead width at stated checkpoints, if the process measures it
metadata: fixture_id fixture_F5, operator_id operator_B09 (pseudonymous), ambient_conditions
sync: shared clock or sync event stated per recording
excluded: robot joint states, contact forces
rights: training and evaluation use; customer part designs stated separately

For arbitrary planning, 2 fixtures, 3 operators and 10 repetitions would give 60 recordings. These inputs are arbitrary, not minimums.

Limits: what this data cannot support

Dispensing video does not contain the tool's commands, the nozzle's pose in a robot frame or the flow rate. A humanoid needs robot, teleoperation or instrumented handheld data to learn where its own hand and tool must be; footage, written procedures, quality records and recognition labels alone do not train control. See human demonstrations vs teleoperated robot demonstrations for the trade-offs.

Further limits:

• Cure and time effects. A bead that looks right at the moment of application may fail later; outcome labels must come from the stage the process cares about.
• Glare and reflections. Wet material on glossy or metallic parts defeats many cameras.
• Fixed-automation bias. If most existing data is from automated cells, human hand technique is absent.
• Single-material coverage. Results from one adhesive do not transfer to another.

Permissions and provenance

Part designs, joint geometries and material specifications may belong to customers or suppliers. Operator hands and faces may be visible, and recordings from process-monitoring cameras may have been taken for other purposes; see process monitoring camera footage. Redaction or removing identifiers does not settle permissions, and masking the bead area would remove the signal you need. Ask who can grant rights over the recordings, the part designs and the material data, and what the operators were told. This is not legal advice.

Questions to ask suppliers

1. Which materials, nozzle sizes and joint types are covered, and are they named in the metadata?
2. Is there a close view of the bead as it forms, or only a wide view?
3. Does the dispenser or controller log pressure, trigger state or flow, and can it be exported with timestamps?
4. How was bead quality judged, by whom and against what limits?
5. Are failed beads, purges and restarts included?
6. Are recordings from hand-held dispensing, automated cells or both?
7. Do any part geometries or materials need to be excluded?
8. Could the plant host a new recording with tool tracking and a specified viewpoint?

Discuss your requirements

If dispensing or sealing is on your task list, describe the materials, 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 the license and delivery. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match. For context, see the humanoid training data guide, the manufacturing hub and the neighboring pages on sanding, polishing and finishing, fixturing and part orientation and battery module assembly.

Describe the workplace demonstration data your robotics project needs

Sources

• Diffusion Policy: Visuomotor Policy Learning via Action Diffusion — policy form that needs observation-action training data; method reference only.