Self-Supervised Learning of Dynamic Planar Manipulation of Free-End Cables

TL;DR

The paper addresses dynamic planar manipulation of free-end cables by learning a forward dynamics model in a tuned simulator and then planning trajectories to place the cable endpoint at a target on the plane. It introduces a self-supervised pipeline with reset-based data collection, a PyBullet-based dynamic cable simulator tuned via differential evolution, and a forward model trained on both simulated and real data to predict endpoint outcomes. Physical experiments on a UR5 with three cables show the approach achieving a median endpoint error of about $22$–$34\%$ of the cable length, outperforming an analytic baseline and a Gaussian Process baseline, and demonstrating effective sim-to-real transfer though with cable-specific challenges (e.g., magnet end). The work provides a practical, data-driven method for dynamic cable manipulation and outlines future work in dynamics randomization and rapid adaptation to new cables, with clear implications for automated cable management in homes and warehouses.

Abstract

Dynamic manipulation of free-end cables has applications for cable management in homes, warehouses and manufacturing plants. We present a supervised learning approach for dynamic manipulation of free-end cables, focusing on the problem of getting the cable endpoint to a designated target position, which may lie outside the reachable workspace of the robot end effector. We present a simulator, tune it to closely match experiments with physical cables, and then collect training data for learning dynamic cable manipulation. We evaluate with 3 cables and a physical UR5 robot. Results over 32x5 trials on 3 cables suggest that a physical UR5 robot can attain a median error distance ranging from 22% to 35% of the cable length among cables, outperforming an analytic baseline by 21% and a Gaussian Process baseline by 7% with lower interquartile range (IQR).

Figures (8)

• Figure 1: Examples of dynamic cable manipulation trajectories with a UR5 robot. After a reset (Section \ref{['ssec:resets']}), the robot performs a dynamic planar action (Section \ref{['ssec:action']}) with two waypoints. The top two rows show a frame-by-frame frontal view of the UR5 as it pulls towards one waypoint (A,B,C) and then towards the second (D,E,F) to get the cable endpoint to reach a desired target. The bottom image shows a top-down view of the setup. There are 32 target positions on the workspace plane marked with small tape, which we use to evaluate performance. The overlaid grid lines are spaced 0.5m apart. We additionally overlay a successful trial repeated 5 times, marked by the star target (lower right), and a failure trial repeated 5 times, marked by the red cross (upper left).
• Figure 2: The process flow for learning dynamic planar cable manipulation (see Section \ref{['sec:method']}). We find a reset motion (a) that consistently brings the cable to a fixed, reset state. We design and parameterize trajectories (b) for dynamically moving the cable endpoint, and execute these trajectories in real (c) to generate $\mathcal{D}_{\rm tune}$, which is then used to tune the dynamic simulator (d). We compare performance of system parameters to select those that maximize coverage and repeatability (e), and use these fixed parameters to generate $\mathcal{D}_{\rm sim}$ (f) and $\mathcal{D}_{\rm real}$. We learn $f_{\rm forw}$ using $\mathcal{D}_{\rm sim}$ along with fine-tuning on $\mathcal{D}_{\rm real}$. The $f_{\rm forw}$ is evaluated on real targets with a physical robot.
• Figure 3: Three cables used in physical experiments, with two polyester ropes and one braided iron wire. Each endpoint has an attached mass, either hooks (first two) or a magnet (third cable). From top to bottom, the cable lengths are 0.62m, 0.65m, and 0.67m without the attachment, and 0.67m, 0.68m, and 0.68m with the attachment. The diameters are 0.011m, 0.002m, and 0.007m.
• Figure 4: Overview of the reset procedure with the UR5, which occurs at the start of each dynamic manipulation. The photos show a frame-by-frame side view of the UR5 as it resets by letting the cable settle (A,B), pulling back (C), casting forward (D,E), and dragging back slightly (F). See Section \ref{['ssec:resets']} for more details. After this reset, the robot executes planar actions, as described in Section \ref{['ssec:action']}.
• Figure 5: Example of two sets of splines that are traced by the end effector during the trajectory (Equation \ref{['eq:act2_wrist']}) of the UR5, for two different $r_0$ values: $r_0 \in \{0.6, 2.0\}$ (not drawn to scale). The reset procedure brings the end effector to $(r_0, \theta_0)$ in the polar space before the start of each dynamic motion. Decreasing $r_0$ increases the curvature of the trajectory.