Short answer
Connector mating and precision insertion are contact-rich tasks, so video alone is weak evidence for a robot. Useful data pairs synchronized video with force or torque, robot or tool pose, and outcome labels, and records tolerances and connector types. Manufacturers can supply context and fixtures; the force and action channels usually require a robot, a teleoperation rig or instrumented capture.
Why insertion data is different
In a peg-in-hole or connector task, success depends on small errors that video cannot resolve: a few tenths of a millimeter of misalignment, a slight tilt, a latch that must be felt rather than seen. A person mating a board-to-board connector uses touch and the sound or feel of a click. A camera sees fingers and a part, often with the contact point hidden by the hand.
That is why insertion appears in robot-learning datasets that record more than images. RH20T, for example, collected more than 110,000 contact-rich real-world manipulation sequences with visual, force, audio and action data per sequence, and paired each robot sequence with a human demonstration video and a language description (Fang et al., 2023). The pairing shows a pattern buyers can reuse: human video as a cheap companion, force and action channels from the robot as the control signal. That is a description of one study's design, not a size guideline.
Policy learning methods then consume such demonstrations. Diffusion Policy frames visuomotor policy as a conditional denoising diffusion process and reports a 46.9% average improvement over prior state of the art across 12 tasks from 4 benchmarks (Chi et al., 2023). Whether a method of that type suits your insertion task depends on your action space, control rate and sensing, which is for your team to establish.
Task anatomy
A typical connector mating cycle:
- Acquire. Pick the connector, cable end or housing from a tray, fixture or the hand of another worker.
- Orient. Rotate to match keying, polarity or pin one.
- Approach. Move to the mating face with the alignment tolerance of the connector family.
- Search or compliant contact. Touch, then wiggle or spiral to find the opening.
- Insert. Push along the axis until the stop, often with a click or latch.
- Verify. Tug test, visual check of seating, or electrical continuity.
- Recover. Back out and retry on jam, or reject the part.
Failure modes include jams on misaligned entry, bent pins, partial seating, wrong keying, cable strain pulling the housing off axis, and connector variants whose insertion force differs from nominal. Frequent variants matter for training: a dataset with only one connector family teaches little about another.
Which data helps each part
| Task part | Helpful channels | Source of the channel |
|---|---|---|
| Acquire and orient | Overhead and wrist-view video; part ID | Human or robot demonstration |
| Approach | End-effector or hand pose over time | Robot state, motion capture or tracked hands |
| Contact and search | Force/torque at wrist or tool; audio | Robot sensor, instrumented tool or glove |
| Insert and latch | Force profile with peak and click signature | Instrumented capture |
| Verify | Test result, continuity or torque record | Manufacturing test station |
| Recover | Annotated failures with cause | Annotation of recorded attempts |
The last row is where existing manufacturing records help. Test-station results and quality logs give outcome labels. See manual electronics assembly, wire harness assembly and snap-fit and press-fit assembly for neighboring tasks.
What each sourcing route can provide
Existing assets. Electronics and automotive manufacturers may hold workstation video, work instructions, fixture drawings, connector part masters and test records. These provide task structure, outcome labels and context. They do not provide robot force or action data, and human video of fingers inserting a connector hides the contact.
Commissioned human demonstrations. A designed recording can use multiple cameras, wrist or head views, tracked hands and instrumented tools. It can include force data if the tool, glove or fixture is instrumented. Sensors alter how the worker handles the part, so ask how that was assessed. The human demonstrations versus teleoperation page lays out the trade-offs.
Robot or teleoperated collection. For a policy that outputs robot actions, you need demonstrations on a robot of the same kind: joint or end-effector actions at a stated control rate, wrist force, and calibrated cameras. This is typically done by specialist collectors with robot cells.
What the data cannot support
Business footage and test records alone cannot train a humanoid to insert connectors. They have no action space, no contact force and often no view of the mating face. Torque records from a screwdriver or press give one channel for one process, not the force profile of a hand pushing a connector. A policy trained on one robot's gripper may not transfer to another's fingers. Insertion tolerance depends on the exact part, and parts differ by supplier lot.
Illustrative specification for an insertion collection
# Illustrative only - synthetic values the buyer would set
task: mate_board_to_board_connector
connector_families: [family_A (keyed), family_B (latching), family_C (friction)]
sensing:
force_torque: wrist 6-axis, sample rate set by buyer, timestamps on shared clock
cameras: 2 fixed + 1 wrist, calibrated intrinsics and extrinsics
actions: end-effector pose commands, control rate stated
labels: step_segments, contact_start, seated_flag, jam_flag, outcome
tolerance_record: nominal clearance and measured offset per trial
variation: lot_id, fixture_id, operator_A12
splits: held-out connector family and held-out lot for evaluation
Limitations and additional preparation
Synchronize every channel. Force, video and pose streams recorded on different clocks will not line up at the instant of contact; see the sensor sync notes in the demonstration acceptance guidance. Calibrate cameras and record hand-eye transforms if a robot is used. Plan for failure cases, because a dataset of only clean insertions teaches little about recovery. Production lines will not stop for extended recording, so expect a dedicated cell or off-shift time.
Choosing the right evidence for your task
Match the data to the question you are asking. If you want to know whether a model can recognize connector families and their keying, images and part masters may be enough. If you want to know whether a policy can complete a mating under positional error, you need demonstrations with actions and contact signals, and an evaluation protocol with deliberate offsets. If you want to assess human technique as a reference, instrumented human demonstrations help, but treat them as a reference for steps and timing, not as robot actions.
Record tolerance explicitly. A dataset that notes initial misalignment per trial lets you stratify results by difficulty. One that does not will average easy and hard insertions into a single number.
Questions to ask suppliers
- Which connector families, pitches and insertion forces are covered, and how many trials per family?
- Is force or torque recorded, from which sensor, at what rate, on which clock?
- Are hands or the contact point visible from any camera?
- What labels mark contact start, seating and failure?
- Are failed and recovered insertions included?
- Was the operator wearing sensors, and how did that change the technique?
- Are test-station results available as outcome labels, and how are they joined to trials?
- Are connector lots and fixture IDs recorded?
- Can a robot-based collection be arranged at the site or a lab, with calibration files delivered?
Discuss your requirements
If you need insertion demonstrations with force or outcome records, describe the connector types, 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. Specialist capture partners can be engaged case by case, and a request does not guarantee a match; SourceX follows up about potential sources. For the broader map, start at the humanoid training data guide or the manufacturing and assembly hub. The screw driving and torque controller data page covers a related process signal.
Discuss your humanoid training-data requirements with SourceX
Sources
- RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot — contact-rich sequences with force, audio, action and paired human video.
- Diffusion Policy: Visuomotor Policy Learning via Action Diffusion — diffusion-based visuomotor policy learning evaluated on 12 tasks.