Robotic wire harness insertion: mechanisms, benefits and trial choices for two approaches

One approach optimizes searching; the other constrains terminal orientation

Kienle and colleagues address connector-to-socket mating. Engineers first write a force-controlled search program. Images, motion and force data collected during execution are used to train a prediction model, which then supports optimization of the approach position and search parameters. The robot still executes the original program with new parameters. Method and authors’ demonstration

This offers a way to retain an existing robot program while reducing search overhead. For a team preparing to reproduce the method, the key investments are a repeatable baseline program, traceable data collection and a structure suitable for gripping. Model training does not resolve material feeding, fixture design or recovery from faults across the station.

Choi and colleagues address terminal insertion into housings at both ends of a single-row flat ribbon cable: guided wire separation, contact-sliding alignment, rocking advancement and clamping to lock the terminals in place. Their small-sample pre-insertion comparison achieved 3/20 with parallel advancement and 18/20 with contact sliding. A separate insertion trial achieved 49/50 with the terminals already aligned, while the complete process achieved 67/80. The hardware remains a 3D-printed prototype that needs adaptation to the product. Paper §III, Tables I–II and §V

This table presents MTTJ’s trial-selection guidance based on the two studies. Their workpieces and experiments differ, so their success rates should not be used to rank the approaches.

Why the methods may work

The studies change two different things: search parameters and the geometry of terminal contact. Understanding this distinction helps identify the preparations your station needs.

Connector mating: reducing unnecessary search motion in the existing program

1. Working force-controlled search program→
2. Collect images, program parameters, motion and force data→
3. Train a process prediction model→
4. Optimize approach and search parameters→
5. Execute new parameters with the original robot program

The original program approaches, probes and retracts at candidate positions to find the insertion position progressively. The prediction model helps the optimizer compare process behavior under different parameters. The paper combines shorter cycles and lower failure probability within the same parameter optimization. Training follows the collection of execution data, and the robot program still executes the optimized parameters. This structure makes the existing process the entry point for the data work. Method §III-A–C

To assess transfer to your station, first check two interfaces: can images, program parameters and action records be associated with the same attempt, and can the optimized parameters be fed back into your control program? If either interface is missing, resolve recording and control integration first. Having a camera and a robot alone is insufficient to estimate the work needed to reproduce the method.

Terminal insertion: using contact to establish an alignment reference

1. Guided wire separation→
2. Housing contacts terminals and slides into alignment→
3. Rocking advancement→
4. Clamping to lock terminals in place

The contact-sliding step first establishes mechanical contact between the housing and terminal end faces. Sliding and rotation then guide the terminals toward the openings. This addresses differences in orientation across a row of terminals, which can make direct parallel advancement collide with the opening edges. Mechanism and pre-insertion steps §III-B

When changing products, terminal length, hole pitch and housing-entry geometry become fixture-evaluation inputs. A team preparing to introduce the method should first check whether the terminals make contact and follow the guides as intended, then address insertion force and lock verification. The prototype mechanism should not be treated as a universal fixture for every terminal.

These flows illustrate principles compiled from the papers. They omit equipment implementation details and are not wiring, manufacturing or robot operating instructions.

Connector mating: how much improvement, and under what conditions?

The following table reorganizes the figures in Kienle and colleagues’ Table III. Each connector underwent 100 tests. Test conditions were not used for training but came from the same distribution as the training data. The comparison was the tactile search program before optimization. The experiments used a UR5e, a camera and custom grippers, with gripping ribs added to the connectors. Training data collection involved about 4,000 attempts per connector. Station-position perturbations were ±2 mm, or ±1.5 mm for B. Paper §IV, §V-B and Table III

The benefit of the same method depends on the original difficulty. C gained 45 percentage points in success rate; D gained 5 percentage points. These differences are calculated by subtraction from the table. Before starting a project, measure your own baseline: does searching account for most of the time, and do failures arise from position uncertainty? If most losses occur in gripping, feeding or latch confirmation, the whole-station benefit of search-parameter optimization requires separate validation.

The paper’s cycle times also cannot directly replace your complete-assembly processing time. An internal comparison should include feeding, retries, manual recovery and time spent handling rejected parts within the same measurement boundary. Timing only successful insertion omits some of the most time-consuming work on the shop floor.

Which stage does 98% describe?

The terminal-insertion study’s 49/50 comes from a separate sample under the condition of insertion after alignment. The 67/80 covers the complete process defined in the paper, with a reported 95% confidence interval of 74.16%–90.25%. The figures answer different questions. Paper Table II and §IV-B

When evaluating a supplier demonstration, ask directly: “At which step does an attempt begin?” and “Are failures before insertion included in the denominator?” Do not subtract results from two different trial sets to locate losses, or multiply them to generate an entire-line result that the paper did not report. To locate losses at your own station, assign the same attempt ID to a workpiece’s feeding, alignment, insertion, locking and inspection stages.

How a station review can determine the next step

The following hypothetical review demonstrates how to select an approach. It does not describe an actual MTTJ or customer station: a team can already mate connectors, but production staff consider the cycle slow and are considering model-based optimization.

First divide a complete assembly attempt into feeding, gripping, searching, insertion, inspection and recovery. Record times and failure locations under the same attempt ID. Then choose according to the records actually obtained:

The review should yield a next trial supported by evidence and identify work not yet worth investing in. Even after selecting a research approach, the paper’s data cannot directly become a return-on-investment estimate for that station. Complete station-cycle data, engineering investment and quality results are still needed.

Turn the research into an executable trial

These preparations are a proposed trial design, not a record of reproducing the papers. The engineers responsible for the equipment and product should determine station safety parameters, gripping force and contact limits for the actual system.

Downloadable records: include retries and manual recovery

The trial-record toolkit contains a blank trial sheet, six demonstration records and field and calculation instructions. Open the CSV files in Excel or LibreOffice. Each row represents one wire harness’s attempt through the complete process; retries remain in the same row. Use a separate batch for trials of individual stages.

The following figures come only from the fictional demonstration data supplied with the article. They illustrate calculations and are not results from the papers or MTTJ testing. Of six attempts, four pass on the first attempt, one passes after a retry, and one fails and requires manual recovery. Total occupied time is 240 seconds. This gives:

All six demonstration records have complete outcome judgments, so these proportions can be calculated. When actual records are incomplete, report known outcomes and the number of missing records first. Do not fill blanks with zero or present the subset with known results as the yield of all attempts. Summarize batches separately when methods, workpieces or timing boundaries differ.

What is available now, and what evidence is worth waiting for?

The mating study provides a paper and the authors’ demonstration page. The pages read for this article did not provide access to a complete code, training-data and fixture-manufacturing package for direct reproduction of the experiment. Confirm availability with the authors or implementer before adoption. The terminal-insertion study provides the mechanism principles and prototype trials, but the same search scope did not yield a complete manufacturing and control package. Neither study supports a commitment on shop-floor introduction time or complete-station pricing.

Useful evidence to follow includes the changes needed for new products, public implementation materials, independent reproductions and longer operating records that include fault recovery. Save your own trial baseline first. When this new evidence appears, you will have a basis for reassessing the investment.

To prepare product conditions, read the flat ribbon cable assembly design guide. To define trial acceptance, read cable assembly testing methods.

Source scope: Kienle et al., arXiv:2503.09409v1 (2025-03-12); Choi et al., arXiv:2608.06996v1 (2026-08-07); authors’ demonstration page. Accessed 2026-09-23. This article studies these public versions without assuming peer review or production validation. MTTJ used AI for research reading and editing and has not physically reproduced the methods described.