Snap-Fit and Press-Fit Assembly: When Visual Data Falls Short

The specification problem

A clip seating into a housing, a bearing pressed onto a shaft or a connector latching are tasks where the visible motion is a small part of what determines success. The operator feels resistance rise, senses the snap, adjusts angle on a slight bind and stops pushing when the part seats. A camera sees a hand moving toward a part and then away.

For a buyer, the consequence is a procurement decision: which sensing to require, who can capture it, and what the resulting data still cannot do. This page helps a robot-learning lead write that requirement. For the simpler question of how visual demonstrations relate to robot learning in general, see what human demonstration video can and cannot do.

What video cannot see

Force-aware datasets exist because of this gap. RH20T recorded visual, force, audio and action data for each of more than 110,000 contact-rich real-world manipulation sequences, and paired each robot sequence with a human demonstration video and a language description (Fang et al., 2023). That design shows what a contact-rich collection usually records, not what an assembly task requires. The sensing a task needs depends on the task, the robot and the learning method.

For background on the signals, see force-torque data and tactile sensing data.

Capture options and their limits

Each route trades fidelity, scale and cost differently.

• Existing station video. Cheapest (see assembly demonstration video), but contains no force, and close-up views of the engagement are rare. Useful for step recognition and as context.
• Instrumented tools or fixtures. A force-torque sensor between tool and part, or under a fixture, records the insertion profile. It describes the interaction, but a sensor in one position cannot tell you about the human hand's contribution.
• Handheld capture devices. A gripper or tool that a person holds, carrying cameras and sensors, produces observations closer to a robot's; see handheld gripper demonstrations.
• Hand motion capture. DexCap is a portable rig that tracks hand motion with SLAM and electromagnetic-field tracking and records 3D scene observations; it was used to learn dexterous policies on six tasks. Its power limit allowed roughly 40-minute sessions, and it has no force sensing (Wang et al., 2024). Motion without force cannot show insertion effort; see motion capture for humanoid motion.
• Teleoperated robot demonstrations. Produce actions in the robot's space, sometimes with force feedback depending on the setup; see teleoperation data collection requirements.

None of these is a default answer. A team with a robot hand and force sensing in hand will want teleoperation data. A team studying task structure may be fine with instrumented human demonstrations.

Decision criteria

Ask these in order:

1. What defines success in the task? If it is a measurable force window or seating depth, the dataset needs that measurement.
2. What will the robot sense? Match the dataset's sensors to the robot's; force data from a sensor the robot lacks is difficult to use.
3. In which action space will the policy act? Human demonstrations need retargeting; see action space and control frequency.
4. How many part variants and tolerances matter? Tight tolerance families need coverage that a single part type cannot give.
5. Is failure data needed? Jams, partial seats and broken clips are rare in routine production and must be staged or logged deliberately.

Illustrative force-aware assembly specification

# Illustrative only - synthetic values; the buyer sets every threshold
task: clip_into_housing (synthetic part IDs housing_H02, clip_K11)
streams:
  video: [engagement_close_up, workstation_wide], 60 fps, constant
  force_torque: 6-axis, fixture-side, 1000 Hz, own hardware clock
  audio: mono microphone near the engagement
  tactile: optional fingertip sensor (state model and calibration)
sync: sync event at episode start; residual drift reported per session
calibration: sensor frame, fixture frame and camera extrinsics, per session
labels:
  phases: approach / align / press / seat / release
  outcome: seated / partial / jammed / broken
  seat_check: independent post-assembly check, method stated
acceptance_examples:
  force_log_gaps: none above buyer-set limit
  episodes_with_outcome_label: 100% (buyer sets the target)
excluded: robot actions unless a robot or handheld device was used

Limitations and additional preparation

Force traces from human demonstrations describe a human strategy applied through a human hand. They need retargeting, or a robot recapture, before they say what a robot should do. Sensor mass and cabling can change how operators work, so ask for evidence of adaptation time. Each session needs calibration and a sync check; see calibration requirements and sensor synchronization requirements. Small part variety limits generalization, and clean, well-seated examples can dominate unless failures are collected. For electronics with delicate connectors, see connector mating and precision insertion.

Permissions and provenance

Instrumented capture is a new collection, so consent can be designed in, but it must cover force and motion data that characterize an individual's technique. Part geometry and latch designs may be customer property. Confirm site access, operator consent and ownership of fixtures and parts. This is not legal advice.

Reviewing a delivery

Spot-check a handful of episodes before accepting a batch. Plot the force trace against the video for a clean seat and for a failed one, and confirm that the force event lines up with the visible moment of engagement within the drift the supplier declared. Check that the sensor's zero offset was recorded at the start of each session, since uncorrected drift makes forces from different days difficult to compare. Confirm that the fixture frame in the calibration file matches the physical fixture used, and that part variants are labeled per episode rather than per session. Finally, check how many episodes end in jams or partial seats; if none do, the set records only the successes, which limits any policy trained or evaluated on it.

Questions to ask suppliers

1. Which sensing is available: force-torque, tactile, audio, motion capture, or only video?
2. Where is each sensor mounted, and what does that position measure?
3. What sampling rates and clock sources are used, and how is sync verified?
4. How does instrumentation affect the operator's technique, and how was that assessed?
5. Can failure cases (partial seats, jams, breakage) be staged and labeled?
6. How is seating verified independently of the operator's judgment?
7. How many part variants and tolerance conditions are included?
8. What calibration records are delivered per session?
9. Who owns the parts and fixtures, and what permissions cover recording them?

Discuss your requirements

If your assembly task depends on force, describe the parts, sensing and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, 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; it can also engage specialist capture partners, scoped case by case. A request does not guarantee a match. See the humanoid training data guide and the manufacturing hub.

Discuss your humanoid training-data requirements with SourceX

Sources

• RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot — RH20T: 110k+ contact-rich sequences with visual, force, audio and action data, paired human video.
• DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation — DexCap portable hand mocap: SLAM and electromagnetic tracking, 6 tasks, about 40-minute sessions, no force sensing.