Online Morphological Adaptation for Tactile Sensing Augmentation

Josie Hughes, Luca Scimeca, Perla Maiolino, Fumiya IidaFrontiers Robotics AI 2021 · DOI · OpenReview

Introduction

In the field of robotics, tactile sensors have traditionally been designed with a fixed morphology—their shape, stiffness, and structural properties are determined during the manufacturing process. However, a static sensor shape is often suboptimal across diverse tasks; a configuration ideal for detecting flat surfaces may fail when interacting with complex geometries or varying textures.

Inspired by biological systems, such as human fingertips that utilize varying compliance to optimize information gain, this research introduces a method for online morphological adaptation. By allowing a sensor to physically restructure itself to "fit" a specific task, the quality of tactile data is significantly enhanced before it reaches machine learning classifiers.

Hardware and Mechanism

The core of this innovation is a morphologically adaptive jamming filter integrated with a standard 50-taxel capacitive tactile sensor. The filter consists of a soft pouch made from latex and silicone, filled with granular particles—specifically, coffee grounds.

Design and operation of the morphologically adaptive jamming filter, showing the vacuum-based mechanism for shape retention.

The system operates via a vacuum-based mechanism. When the vacuum is off (unjammed state), the filter is soft and deformable, allowing it to be molded against a template or an object. Once the vacuum is applied (jammed state), the particles lock together, effectively "freezing" the filter into a specific shape. This allows a UR5 robot arm to perform a two-phase interaction: first molding the sensor to a desired template, and then performing controlled tactile exploration, such as a rubbing motion, against a test object.

Experimental Methodology

The system was evaluated using two distinct datasets to test its versatility:

  1. Feature-Based Set: Eight 3D-printed shapes designed to isolate tactile features like stiffness (hard vs. soft), roughness (smooth vs. ridged), and geometry (square vs. round).
  2. Real-World Set: Seventeen everyday items, ranging from pens and batteries to organic objects like an orange and a seashell.

Data processing involved reducing high-dimensional tactile inputs (12,500 dimensions per trial) using Principal Component Analysis (PCA). The resulting data was then evaluated across nine standard machine learning classifiers, including Support Vector Machines (SVM), Random Forests, and Gaussian Processes.

Results and Discussion

The implementation of the adaptive filter yielded substantial improvements in object classification. The research demonstrated that physical pre-processing—optimizing the sensor's structure to capture relevant information—is often more impactful than increasing the complexity of the downstream machine learning models.

Quantitative comparison showing the performance boost provided by adaptive and shaped filters over fixed morphologies.

Key findings include:

  • Accuracy Gains: The adaptive filter increased classification accuracy by up to 28% compared to non-adaptive, fixed-shape sensors.
  • The "Lock and Key" Effect: When a filter is molded to a specific object, it creates a physical fit that amplifies subtle differences between similar objects in the sensor's latent space.
  • Information Density: PCA revealed that adaptive and ridged filters could retain over 60% of data variance in a single dimension, whereas soft, non-adaptive filters produced cluttered, inseparable data.
  • Generalization: On the 17-object real-world set, the adaptive filter was the top performer for the majority of the classifiers tested.

Conclusion and Future Work

This work demonstrates that online morphological adaptation is a powerful tool for augmenting robotic tactile sensing. By treating the sensor's physical shape as a programmable component, robots can achieve higher discrimination accuracy with simpler computational models.

While current limitations include a bottleneck in spatial resolution (approximately 3mm) and the time required for manual molding, future iterations may utilize automated molding stations or evolutionary algorithms to autonomously determine the optimal sensor shape for unknown objects in real-time.

Summary updated

Abstract

Sensor morphology and structure has the ability to significantly aid and improve tactile sensing capabilities, through mechanisms such as improved sensitivity or morphological computation. However, different tactile tasks require different morphologies posing a challenge as to how to best design sensors, and also how to enable sensor morphology to be varied. We introduce a jamming filter which, when placed over a tactile sensor, allows the filter to be shaped and molded online, thus varying the sensor structure. We demonstrate how this is beneficial for sensory tasks analyzing how the change in sensor structure varies the information that is gained using the sensor. Moreover, we show that appropriate morphology can significantly influence discrimination, and observe how the selection of an appropriate filter can increase the object classification accuracy when using standard classifiers by up to 28%.
BibTeX
@article{hughes2021online,
  author = {Josie Hughes and Luca Scimeca and Perla Maiolino and Fumiya Iida},
  title = {{Online Morphological Adaptation for Tactile Sensing Augmentation}},
  journal = {Frontiers Robotics AI},
  year = {2021},
  doi = {10.3389/frobt.2021.665030},
}
Figure