Extremum Seeking Controlled Wiggling for Tactile Insertion

TL;DR

The paper tackles robotic insertion tasks with complex geometries by emulating a human-like wiggle strategy guided by tactile feedback. It introduces a model-free Extremum Seeking Control framework that wiggles a 6-DOF end-effector pose while monitoring strain via GelSight Mini sensors to minimize contact effort and maximize insertion depth. The approach is evaluated on four locks and five AutoMate benchmark assemblies, outperforming CMA-ES baselines and demonstrating strong generalization, including vision-based initialization on unseen objects. Results show high success rates (e.g., up to 98% on certain locks) and robust performance across varied initial poses without per-object tuning or detailed contact models. This work presents a practical, learning-free paradigm for tactile insertion with potential applicability to diverse robotic assembly tasks.

Abstract

When humans perform complex insertion tasks such as pushing a cup into a cupboard, routing a cable, or putting a key in a lock, they wiggle the object and adapt the process through tactile feedback. A similar robotic approach has not been developed. We study an extremum seeking control law that wiggles end effector pose to maximize insertion depth while minimizing strain measured by a GelSight Mini sensor. Evaluation is conducted on four keys featuring complex geometry and five assembly tasks featuring basic geometry. On keys, the algorithm achieves 71% success rate over 120 trials with 6-DOF perturbations, 84% over 240 trials with 1-DOF perturbations, and 75% over 40 trials initialized with vision. It significantly outperforms a baseline optimizer, CMA-ES, that replaces wiggling with random sampling. When tested on a state-of-the-art assembly benchmark featuring basic geometry, it achieves 98% over 50 vision-initialized trials. The benchmark's most similar baseline, which was trained on the objects, achieved 86%. These results, realized without contact modeling or learning, show that closed loop wiggling based on tactile feedback is a robust paradigm for robotic insertion.

Figures (6)

• Figure 1: By wiggling the 6 degree of freedom pose of an object grasped between two GelSight Mini tactile sensors and observing a strain-like quantity through the optical flow in the GelSight cameras, an extremum seeking control law performs insertion. All parameters are sinusoidally modulated simultaneously but at different frequencies, allowing for the estimation of a direction that minimizes strain and maximizes insertion depth along the Y axis.
• Figure 2: The extremum seeking controlled pipeline for wiggling-based tactile insertion. The instantaneous parameters $\theta$ control the pose of the tip of an object through a UR10 robot arm. The strain that the object exerts on the GelSight Mini's gel pad is observed via a displacement of the corners of a tracked patch in the sensor image feed, $L_{strain}$. The objective to be minimized is the sum of $L_{strain}$ plus $L_{insertion}$, where $L_{insertion}$ represents the depth of insertion into the lock. The extremum seeking control seeks to minimize the objective by adjusting the parameter estimates $\hat{\theta}$. As is standard in Extremum Seeking Control, $\theta$ is a modulated version of $\hat{\theta}$ with each parameter modulated at a different frequency. The high pass filter removes the DC component from the objective signal, demodulation determines the slope of the objective's gradient, and the low pass filter averages the feedback signal with greater high-frequency attenuation than the integrator.
• Figure 3: A strain-like measurement is measured directly from the images returned by the GelSight Mini sensor. Each incoming frame is iteratively registered with a Lucas-Kanade style homography estimator to the first frame. The tracked patch has 10% margins with respect to the full frame and is exemplified by the area within the white box with red corners. The Euclidian norm of the corner displacements in pixels from their original location is used as the strain-like quantity $L_{strain}$.
• Figure 4: The loss and estimated parameters (end-effector pose) converge as the key is inserted. The Y parameter increases steadily because it corresponds to the insertion axis. The trial pictured was initialized with $[1.1, 0.0]$ millimeters of translation along the X, Z axis and $[3.4, -7.4, 5.7]$ degrees of rotation about the X, Y, Z axis.
• Figure 5: The four types of key and lock pairs tested. L1 is a common pin-tumbler lock used on front doors. L2 is a dimpled cam lock that uses pin-tumblers like L1, but they also press on the sides of the key into the dimples. L3 is a tubular lock with a circular shape that must be pressed into the lock's circular opening. L4 is a disc-detainer which features rotating discs inside the lock that must be aligned.