Behavioral Cloning for Robotic Connector Assembly: An Empirical Study

Andreas Kernbach; Daniel Bargmann; Werner Kraus; Marco F. Huber

cs.ROFeb 25, 2026

Behavioral Cloning for Robotic Connector Assembly: An Empirical Study

Authors: Andreas Kernbach, Daniel Bargmann, Werner Kraus, Marco F. Huber

Organizations: The authors are with the Fraunhofer Institute for Manufacturing Engineering and Automation IPA, Stuttgart, Germany · Institute of Industrial Manufacturing and Management IFF, University of Stuttgart, Germany

Abstract

Automating the assembly of wire harnesses is challenging in automotive, electrical cabinet, and aircraft production, particularly due to deformable cables and a high variance in connector geometries. In addition, connectors must be inserted with limited force to avoid damage, while their poses can vary significantly. While humans can do this task intuitively by combining visual and haptic feedback, programming an industrial robot for such a task in an adaptable manner remains difficult. This work presents an empirical study investigating the suitability of behavioral cloning for learning an action prediction model for connector insertion that fuses force-torque sensing with a fixed position camera. We compare several network architectures and other design choices using a dataset of up to 300 successful human demonstrations collected via teleoperation of a UR5e robot with a SpaceMouse under varying connector poses. The resulting system is then evaluated against five different connector geometries under varying connector poses, achieving an overall insertion success rate of over 90 %.

Figures & tables

Fig. 1: BC predictor based on human demonstrations. Force–torque controller targets are generated using a space mouse and a NN is trained to imitate insertion strategies from camera and force–torque sensor data.

Panel of Figure 2 — Fig. 2: Hardware setup used in the experiments.

Model	$\bm{\mu}_{\mathrm{SR}}$	$\bm{\sigma}_{\mathrm{SR}}$	MSE	$\bm{n}_{\theta}$	$\bm{t}_{\text{Inf, max}}$
RegNetx3_2GF	93.3	5.80	0.0376	14.6	3.14
DenseNet201	91.1	15.4	0.0373	18.7	6.04
DenseNet121	90.0	12.0	0.0370	7.3	4.20
EfficientNetV2S	84.4	24.1	0.0372	20.6	4.45
RegNety8GF	81.1	27.1	0.0374	38.0	2.74
V-ResNetR3D	80.0	31.8	0.0384	33.4	5.99

TABLE I: Comparison of image backbones by success rate (

\mu_{\mathrm{SR}}

\sigma_{\mathrm{SR}}

), MSE , total number of model parameters

n_{\theta}

(in millions) and worst-case inference time

t_{\text{Inf},\max}

(in milliseconds).

Panel of Figure 4 — Fig. 3: Mean success rate over different history lengths $k$ and prediction horizons $h$ on a shared $x$ -axis. In blue different $k$ given $h=1$ . In green different $h$ given $k=10$ .

Model	$\bm{\mu}_{\mathrm{SR}}$	$\bm{\sigma}_{\mathrm{SR}}$	MSE	$\bm{n}_{\theta}$
Feedforward	96.7	2.86	0.0342	14,637,468
LSTM	80.0	5.43	0.0333	14,955,100
ROCKET	80.0	16.3	0.0356	14,752,476
1DCNN	66.7	11.0	0.0334	14,672,604

TABLE II: Comparison of different sensor backbones.

Model	$\bm{\mu}_{\mathrm{SR}}$	$\bm{\sigma}_{\mathrm{SR}}$	MSE	$\bm{n}_{\theta}$
Feedforward	96.7	2.86	0.0342	14,637,468
$\text{Transformer}_{\text{Full}}$	86.7	13.4	0.0388	14,748,966
$\text{Transformer}_{\text{Encoder}}$	80.0	6.65	0.0372	14,830,950

TABLE III: Comparison of different action heads.

Panel of Figure 7 — Fig. 4: Mean success rate given a number of demonstrations $m$ with a comparison of a rule based stride search strategy from [ 3 ] .



Manufacturer	Molex	TE	TE	JAE
Series	Mini-Fit	MQS	AMP MCP	MX19
ID	A	B	C	D
$\mu_{\mathrm{SR}}$	93.3	94.4	90.0	93.3
$\sigma_{\mathrm{SR}}$	5.47	2.01	8.10	5.47

TABLE IV: Evaluation on different connector geometries with success rate (

\mu_{\mathrm{SR}},\sigma_{\mathrm{SR}}

), human/ BC insertion time in seconds (and steps), and translational tolerance in millimeter.

Explore similar work

Touch2Insert: Zero-Shot Peg Insertion by Touching Intersections of Peg and Hole
Mar 4, 2026Masaru Yajima, Yuma Shin, Rei Kawakami +2Robotic Insertion Tactile Sensing for Robotic Manipulation
Institute of Science Tokyo, Tokyo, Japan. · Mitsubishi Electric, Kanagawa, Japan.
InsertAnything: Generalizable Contact-Rich Precision Insertion from Simulation to Reality
Sep 21, 2026Zhenghua Ma, Xinpan Meng, Zeyu Liu +4RL for Robotics Contact-Rich Robotic Manipulation
School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China, and the State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · PaXini AI Technology (Beijing) Co., Ltd., Beijing 100080, China
Generalizable Robotic Insertion with World Models
Sep 23, 2026Nicklas Hansen, Iretiayo Akinola, Yijie Guo +7Zero-Shot Robotic Manipulation World Models for Robotics
NVIDIA · University of California San Diego · Work completed during an internship at NVIDIA +1

Mar 4, 2026J/K move · Enter open · S save

Touch2Insert: Zero-Shot Peg Insertion by Touching Intersections of Peg and Hole

Masaru Yajima, Yuma Shin, Rei Kawakami, Asako Kanezaki+1

Institute of Science Tokyo, Tokyo, Japan. · Mitsubishi Electric, Kanagawa, Japan.

Robotic Insertion Tactile Sensing for Robotic Manipulation Zero-Shot Robotic Manipulation Object Pose Estimation

Open paper AskPDF arXiv

Reliable insertion of industrial connectors remains a central challenge in robotics, requiring sub-millimeter precision under uncertainty and often without full visual access. Vision-based approaches struggle with occlusion and limited generalization, while learning-based policies frequently fail to transfer to unseen geometries. To address these limitations, we leverage tactile sensing, which captures local surface geometry at the point of contact and thus provides reliable information even under occlusion and across novel connector shapes. Building on this capability, we present Touch2Insert, a tactile-based framework for arbitrary peg insertion. Our method reconstructs cross-sectional geometry from high-resolution tactile images and estimates the relative pose of the hole with respect to the peg in a zero-shot manner. By aligning reconstructed shapes through registration, the framework enables insertion from a single contact without task-specific training. To evaluate its performance, we conducted experiments with three diverse connectors in both simulation and real-robot settings. The results indicate that Touch2Insert achieved sub-millimeter pose estimation accuracy for all connectors in simulation, and attained an average success rate of 86.7% on the real robot, thereby confirming the robustness and generalizability of tactile sensing for real-world robotic connector insertion.

Abstract

Explore similar work

Touch2Insert: Zero-Shot Peg Insertion by Touching Intersections of Peg and Hole

InsertAnything: Generalizable Contact-Rich Precision Insertion from Simulation to Reality

Generalizable Robotic Insertion with World Models