AI-based Framework for Robust Model-Based Connector Mating in Robotic Wire Harness Installation

Claudius Kienle, Benjamin Alt, Finn Schneider, Tobias Pertlwieser, Rainer Jäkel, Rania Rayyes · 2025

Read AI-based Framework for Robust Model-Based Connector Mating in Robotic Wire Harness Installation by Claudius Kienle, Benjamin Alt, Finn Schneider, Tobias Pertlwieser, Rainer Jäkel, Rania Rayyes (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-09409.txt.

Introduction

Car wire harnesses still go in by hand, because plugging in their connectors takes a feel that robots lack. Six researchers present an AI framework in 2025 that mates connectors with a robot arm. The system joins force control with a model that learns from sight, touch, and joint position together. It tunes its search-and-insert moves with a fast optimization method. A data pipeline gathers training examples on its own, so engineers need little machine learning skill.

The tuned programs run on standard factory controllers, which lets experts review and approve them like normal robot code. On a center console task, the system cut cycle times and held up better than hand-programmed approaches.

Catalog record

CREATORClaudius Kienle, Benjamin Alt, Finn Schneider, Tobias Pertlwieser, Rainer Jäkel, Rania Rayyes
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-09409

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

AI-based Framework for Robust Model-Based Connector Mating in Robotic Wire Harness Installation Claudius Kienle, Benjamin Alt, Finn Schneider, Tobias Pertlwieser, Rainer Jäkel, Rania Rayyes Abstract Despite the widespread adoption of industrial robots in automotive assembly, wire harness installation remains a largely manual process, as it requires precise and flexible manipulation. To address this challenge, we design a novel AI-based framework that automates cable connector mating by integrating force control with deep visuotactile learning. Our system optimizes search-and-insertion strategies using first-order optimization over a multimodal transformer architecture trained on visual, tactile, and proprioceptive data. Additionally, we design a novel automated data collection and optimization pipeline that minimizes the need for machine learning expertise. The framework optimizes robot programs that run natively on standard industrial controllers, permitting human experts to audit and certify them. Experimental validations on a center console assembly task demonstrate significant improvements in cycle times and robustness compared to conventional robot programming approaches. Videos are available under https://claudius-kienle.github.io/AppMuTT . I INTRODUCTION Fig. 1: Robotic mating of electrical connectors in a center console: Using our learned, neural model of the search-and-insertion strategy, the search pattern (bottom left, red) is optimized (green) to ensure robust installation with minimal temporal overhead. Industrial robots have been central to the automation of assembly tasks in the automotive industry for several decades [ 1 ] . Recent advances in sensor and actor technologies such as cost-effective force-torque sensors and 3D cameras have facilitated software- and data-driven automation, enabling the use of robots in increasingly dynamic, flexible assembly applications [ 2 ] . Despite these advances, the installation of wire harnesses has remained largely confined to manual labor and expensive, specialized mechanical engineering solutions [ 3 ] . This is largely due to the process variances induced by the flexibility of cables, the wide range of product variants, and strict process requirements with respect to cycle time and robustness [ 4 ] . As modern automobiles incorporate more connected sensors and actuators, partly driven by the demands of autonomous driving, the number and complexity of wire harnesses are rapidly increasing. As a result, wire harness installation is becoming a greater bottleneck and cost factor in the automotive value chain [ 4 ] . One of the final steps in wire harness installation in automobiles is the mating of the cable connectors [ 5 ] . This installation step poses considerable challenges for robotic automation, as it requires handling a large variety of connector geometries [ 4 ] and accessing sockets located in the drivetrain, center console, dashboard, and other areas that are hard to reach for 6- dof ( dof ) robot arms [ 6 ] . Moreover, the pre-positioning of the wire harness and installation targets is subject to stochastic variances on the order of several millimeters [ 3 ] . State-of-the-art techniques for robotic connector mating use force-controlled or vision-based search and insertion, inherently facing a tradeoff between robustness and execution time. To address this challenge, we propose a novel framework for the automated mating of wire harness cable connectors, combining advanced machine learning techniques [ 7 ] with a streamlined, automation-friendly process [ 8 ] . It leverages multimodal visuotactile learning and force control to optimize the connector insertion strategy, enhancing both efficiency and robustness. At the core of our framework is a mutt ( mutt ) [ 7 ] , trained on visual, tactile, and proprioceptive data to capture the stochastic characteristics of the insertion process. By applying mutt as a predictor in a model-based optimizer [ 9 ] , we enable the system to adapt its search-and-insertion strategy, leading to improved cycle times and insertion success rates. Our AI-based framework is complemented by a fully automated pipeline for data collection, training, and optimization, minimizing the need for human intervention and making the system accessible to non-experts in AI. We validate our framework in real-world experiments on installing a wire harness into a center console, and evaluate its performance across five connector types with different geometries. Search optimization results in updated parameters on the robot controller, ensuring that the resulting search strategy remains interpretable, auditable, and certifiable by industry professionals. II RELATED WORK II-A Robotic Wire Harness Installation Wire harness installation features several challenging perception and manipulation problems, such as the accurate perception of the position and state of deformation of the harness [ 10 , 11 , 3 , 12 ] , robust grasping of the wires or connectors [ 13 , 14 , 15 ] , outlaying and fixing of the harness in the chassis [ 16 , 6 , 17 ] and mating of the electrical connectors [ 5 , 18 , 19 ] . This paper focuses on the latter task of connector insertion, which remains a manual process in the industrial state of the art [ 12 ] . Researchers have proposed several approaches for robotic connector mating, most of which use a combination of visual and force feedback. Sun et al. [ 19 ] propose to measure the tilt angle and displacement of the grasped connector with two 2D cameras and use a force- and torque-controlled search and insertion strategy to mate the connector with the socket. Di et al. [ 18 ] propose a hybrid force-vision control strategy for online motion control. Yumbla et al. [ 5 ] propose a gripper design that allows grasping the cable so that the connector is aligned and firmly arrested, considerably reducing the uncertainty induced by imprecise grasping. Our framework improves upon previous methods for force and vision-controlled connector mating [ 19 , 20 ] by introducing a gripper-friendly connector design (cf. Fig. 4 ) and an optimization approach that tailors the search strategy to the stochastic variance of the insertion process and the specific characteristics of the current environment. By learning the environment distribution with mutt and adapting the search accordingly, our method minimizes unnecessary exploration, reducing search overhead while enhancing the robustness and efficiency of the mating process. II-B Data-Driven Robot Program Optimization Force-controlled search and insertion strategies compensate for sources of process noise, such as elastic deformation of the socket, that are difficult to detect via purely vision-based approaches. Force-controlled search and careful moment-free insertion trade off robustness for time. Data-driven methods permit to optimize search and insertion strategies to minimize additional search overhead. While a variety of general-purpose robot program optimization approaches have been proposed [ 21 , 22 , 23 , 24 ] , model-based parameter optimization has emerged as a promising approach for peg-in-hole assembly problems, as it avoids trial and error at runtime in favor of optimizing over a learned model of the process [ 25 ] . In prior work, we introduced spi ( spi ), a first-order parameter optimizer operating on a differentiable, partially learned model of a robot program [ 9 , 26 ] . Related approaches have applied optimization by gradient descent over differentiable programs to optimize motion controller parameters [ 27 , 28 , 29 , 30 ] . spi has been applied to peg-in-hole electronics assembly [ 26 ] , but without conditioning on the current environment, limiting optimization to process variance alone. In prior work [ 7 ] , we extended spi with mutt to

Read the complete text