Source note
Wiggle and Go! System Identification for Zero-Shot Dynamic Rope Manipulation
Summary
Wiggle and Go! estimates rope dynamics from one short probing motion, then uses those estimated parameters to plan a dynamic rope action in simulation and execute it in the real world without trial-and-error. The paper targets zero-shot rope tasks where failed attempts can be costly or unsafe.
Problem
- Dynamic rope manipulation is hard because rope behavior depends on hidden properties such as stiffness, damping, mass distribution, and link count.
- Prior methods often need large real-world datasets or several task attempts to adapt, which is a poor fit when a bad throw can tangle, damage, or miss in ways that are hard to recover from.
- The paper aims to infer rope-specific dynamics once, reuse that estimate across tasks, and execute a goal-conditioned action zero-shot.
Approach
- The method has two stages: perform a predefined low-risk "wiggle" to observe rope motion, then predict rope parameters and use them to optimize a task action.
- A temporal convolutional network takes tracked 2D rope keypoints and angle features from the wiggle and predicts 9 simulator parameters, including stiffness, damping, rope length, mass per unit length, lead mass, and link count.
- The model is trained in simulation, where parameter labels are known, with sim-to-real randomization for camera calibration noise, tracking noise, delayed recording, and masked time windows.
- For each task goal, the system runs CMA-ES trajectory optimization in Drake using the predicted rope parameters, then executes the resulting robot trajectory on an xArm 7.
- The same system-identification module is task-agnostic and is reused across three downstream tasks: 3D target striking, lobbing, and draping.
Results
- On real 3D target striking, the method reports 3.55 cm average accuracy when using predicted rope system parameters, compared with 15.34 cm when the action model is not informed by system parameters.
- In the introduction, the paper also reports median striking accuracy of 3.55 cm in real and 2.1 cm in simulation, versus 15.29 cm in real and 12.8 cm in simulation for the non-parameter-informed baseline.
- For transfer of identified dynamics to a different motion context, the paper reports a Pearson correlation coefficient of 0.95 between Fourier frequencies of predicted-rope and real-rope behavior on an unseen trajectory.
- On more complex downstream tasks, the paper claims over 50% success rate for lobbing and draping.
- In the wiggle ablation table, the main wiggle gives low mean absolute error on several predicted parameters, including 0.098 for link count, 0.006 m for rope length, 0.010 N·s/m for ball damping, 0.002 m for rope radius, 0.007 kg/m for mass per unit length, 0.005 kg for lead mass, and 0.111 N/m for ball stiffness, with random wiggles performing worse on several parameters.
- The paper claims one wiggle observation can support multiple manipulation policies without retraining the identification module.