Robots of the Lost Arc: Self-Supervised Learning to Dynamically Manipulate Fixed-Endpoint Cables

TL;DR

This work tackles dynamic manipulation of fixed-endpoint cables to cross obstacles, knock targets, and weave between obstacles using a UR5 robot. It introduces INDy, a self-supervised imitation-learning framework that predicts a 3D apex point which, when combined with a minimum-jerk trajectory via a quadratic program, yields fast, feasible arm motions without explicit cable physics models. The approach is trained with self-collected real data, validated in simulation, and demonstrated on physical hardware across five cables, achieving up to 81.7% vaulting success and substantially higher performance than baselines. Findings indicate strong potential for apex-driven dynamic manipulation and reveal practical considerations for sim-to-real transfer and generalization to cable properties.

Abstract

We explore how high-speed robot arm motions can dynamically manipulate cables to vault over obstacles, knock objects from pedestals, and weave between obstacles. In this paper, we propose a self-supervised learning framework that enables a UR5 robot to perform these three tasks. The framework finds a 3D apex point for the robot arm, which, together with a task-specific trajectory function, defines an arcing motion that dynamically manipulates the cable to perform tasks with varying obstacle and target locations. The trajectory function computes minimum-jerk motions that are constrained to remain within joint limits and to travel through the 3D apex point by repeatedly solving quadratic programs to find the shortest and fastest feasible motion. We experiment with 5 physical cables with different thickness and mass and compare performance against two baselines in which a human chooses the apex point. Results suggest that a baseline with a fixed apex across the three tasks achieves respective success rates of 51.7%, 36.7%, and 15.0%, and a baseline with human-specified, task-specific apex points achieves 66.7%, 56.7%, and 15.0% success rate respectively, while the robot using the learned apex point can achieve success rates of 81.7% in vaulting, 65.0% in knocking, and 60.0% in weaving. Code, data, and supplementary materials are available at https: //sites.google.com/berkeley.edu/dynrope/home.

Figures (6)

• Figure 1: Long exposure photo of a UR5 robot dynamically manipulating an orange cable fixed at one end (far left) to knock the red target cup off the white obstacle using the learned 3d apex point and minimum-jerk robot trajectory.
• Figure 2: Illustration of the 3 tasks: Vaulting, Knocking, and Weaving in real experiments with an orange 18-gauge cable.
• Figure 3: Repeatability without (left) and with (right) taut-pull reset. Both images show an overlay of 20 ending cable configurations after sequentially applying the same left-to-right vaulting trajectory 20 times. Left: without applying resets before the motion, there is high variation in ending configurations. Right: resetting the cable via a taut-pull before the motion results in nearly identical configurations across 20 left-to-right attempts.
• Figure 4: Cable trajectory for data collection in simulation (top) vs. cable trajectory in real (bottom) for the vaulting task after applying the same apex point configuration.
• Figure 5: Left: Using three points in a curve to control the trajectory of the cable. The start and end points are fixed using the arm configurations. We change the trajectory by varying the apex configuration. Right: Design of physical experiments. We randomize the white Lego bricks’ location within the yellow shaded area. For vaulting and knocking, the width of the yellow area is 1.5 m, and 1.0 m for weaving. The yellow area is divided into three difficulty tiers.