Harnessing with Twisting: Single-Arm Deformable Linear Object Manipulation for Industrial Harnessing Task

Xiang Zhang, Hsien-Chung Lin, Yu Zhao, Masayoshi Tomizuka

Introduction

Automating the process of wire harnessing—routing and securing cables into intricate layouts—remains a significant hurdle in industrial robotics. Unlike rigid components, wires are Deformable Linear Objects (DLOs) characterized by infinite degrees of freedom and highly nonlinear dynamics. Traditional approaches to managing these complexities typically involve expensive dual-arm systems or delicate tactile sensors. Dual-arm configurations, while functional, are often too bulky for the confined spaces of industrial electrical cabinets, and tactile sensors frequently suffer from durability issues in high-cycle environments.

This paper presents a robust single-arm robotic pipeline that achieves industrial-grade precision using a standard collaborative robot and force/torque sensing. The core innovation lies in a "twisting" strategy that enables a single gripper to maintain the high tension required for routing and snapping wires into fixtures, effectively replacing the need for a second stabilizing arm.

Methodology

The Twisting Mechanism

The fundamental challenge in single-arm DLO manipulation is maintaining sufficient tension without the wire slipping through the gripper. The researchers discovered that by twisting the wire against a custom-designed, grooved 3D-printed gripper, they could significantly increase the contact area and friction. This "twisting" action acts as a force multiplier, allowing the robot to exert forces exceeding 10 Newtons—the threshold required to snap wires into industrial clamps—whereas a standard grasp would fail at approximately 4 Newtons.

Koopman-Operator Predictive Control

To manage the unpredictable physics of a stretching wire, the system employs a Model Predictive Controller (MPC) based on the Koopman operator. The Koopman operator framework "lifts" the nonlinear dynamics of the DLO into a higher-dimensional linear space. This mathematical transformation allows the controller to perform real-time, accurate predictions of how robot end-effector movements will translate into wire tension, maintaining the computational efficiency of linear control while capturing the accuracy of nonlinear modeling.

System Architecture and Planning

The pipeline is supported by three additional pillars:

  1. Harnessing Motion Planner: A high-level planner that generates waypoints around fixtures to ensure the wire follows the intended path while avoiding environmental collisions.
  2. Insertion Primitives: Pre-defined motion sequences tailored for specific industrial fixtures, such as "C-shaped" clamps (used as routing pivots) and "U-shaped" clamps (used for final locking).
  3. Fix-point Switching: As the wire is successfully secured into a clamp, the system dynamically updates its internal model. The newly engaged clamp becomes the new "fixed origin" for the subsequent segment of the wire, allowing for sequential routing across complex boards.

Experimental Validation

Dataset and Training

The researchers utilized a targeted offline dataset consisting of 40 real-world wire-following trajectories. To enhance the robustness of the Koopman dynamics model, they applied geometric data augmentation—specifically rotating the trajectories—to expand the training set to 400 samples. This approach demonstrates that high-performance DLO manipulation can be achieved without the need for massive, labor-intensive datasets.

Performance Benchmarking

Testing was conducted using a FANUC CRX-10iA collaborative robot on the NIST Assembly Task Board 4, a standard benchmark for industrial assembly. The system was evaluated on its ability to track tension and successfully complete multi-step routing tasks.

Results and Discussion

The "twisting" approach demonstrated clear superiority over traditional linear models and non-twisting baselines:

  • Tension Reliability: The proposed method consistently tracked tension targets up to 10N. In contrast, the "No-twisting" baseline suffered from immediate wire slippage once tension surpassed 4N.
  • Task Success Rates: In single-wire routing trials, the system achieved a 100% success rate across ten trials. In more complex scenarios involving multiple wires—where the risk of entanglement is high—the system maintained an 88.8% (8/9) success rate.
  • Control Stability: Compared to standard linear dynamics models without Koopman lifting, the proposed controller showed significantly higher stability, avoiding the excessive force spikes that typically lead to wire damage or connector pull-out.

By maintaining constant tension, the system effectively simplifies a complex 3D deformation problem into a more manageable 2D geometry problem. This tension-centric approach makes state estimation significantly more robust than methods attempting to track the geometry of a slack or "floppy" wire.

Limitations and Future Work

While the current pipeline is highly effective for industrial applications, certain constraints remain. The system currently assumes the wire is pre-inserted into a starting connector and that the robot has already established an initial grasp. Additionally, the long-term effects of repeated twisting on wire insulation (material fatigue) require further study. Future research will focus on online adaptation techniques, enabling the robot to autonomously learn the physical properties (stiffness, thickness) of unfamiliar wire types in real-time.

Summary updated

BibTeX
@inproceedings{zhang2024harnessing,
  author = {Xiang Zhang and Hsien-Chung Lin and Yu Zhao and Masayoshi Tomizuka},
  title = {{Harnessing with Twisting: Single-Arm Deformable Linear Object Manipulation for Industrial Harnessing Task}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2024},
  doi = {10.1109/iros58592.2024.10802801},
}
Figure