Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

Flexible Piezoelectric Tactile Sensing System for Intelligent Object Recognition

Razan Khalifeh, Mohamad Yaacoub, Christian Gianoglio, Moustafa Saleh, Maurizio Valle, Yahya Abbass

November 2026

TL;DR Two flexible 16-element piezoelectric arrays in silicone caps on a TIAGo gripper recognize six everyday objects from the first 100 ms of a grasp. A small 2D-CNN on spatio-temporal heatmaps reaches 0.99 F1 with all 32 sensors and runs on a microcontroller with about 101 ms end-to-end latency.

What they do

Piezoelectric tactile sensors respond to changing force. They spike when a gripper closes and opens and stay nearly silent during a steady hold. The authors put two 16-element flexible piezoelectric arrays in silicone caps on the fingers of a TIAGo gripper. They then measure how accurately, and how soon after contact, the 32 signals identify which of six everyday objects is being grasped.

How it works

  • Sensor. Each array is a 4x4 grid of screen-printed P(VDF-TrFE) sensors at 5 mm resolution on a flexible PCB, cast into a removable Dragon Skin cap. Sensitivity rises with frequency from 1.5 to 31.96 pC/N, and a 32-channel board samples every sensor at 500 Hz.
  • Data. The robot grasps, holds and releases each object 50 times, with its pose in the gripper varied, for 300 trials. The objects are a glass bottle, a half-full plastic bottle, a TV remote and three 3D-printed YCB fruits.
  • Three classifiers. An SVM on four per-sensor features, a 1D-CNN and a C-RNN on raw signals, and a small 2D-CNN on heatmaps laid out like the sensor grid. Each is tested with windows of 50-200 samples and with the 16, 24 or 32 highest-energy sensors.

Results

  • Short windows work best for every method. The 2D-CNN on 16 sensors falls from 0.9796 F1 at 50 samples to 0.9077 at 200.
  • More sensors help. With 50-sample windows the 2D-CNN reaches 0.9796, 0.9846 and 0.9903 F1 with 16, 24 and 32 sensors, against 0.9661 for the SVM and 0.9618 for the 1D-CNN with all 32.
  • On an STM32 microcontroller the 2D-CNN uses 7.65 kB of flash and runs in 0.88 ms (63.8 µJ per inference). End-to-end latency from contact onset is about 101 ms, mostly spent filling the 100 ms window.

Limitations

As the paper notes, the study covers six objects, one robot and one scripted grasp-hold-release motion. Accuracy is measured offline on recorded trials. The embedded deployment reports only latency and energy, not online recognition accuracy.

Figures

From Figure 1: One sensor and the 4x4 array layout, how the silicone sensing caps are made, the system block diagram,...
From Figure 1: One sensor and the 4x4 array layout, how the silicone sensing caps are made, the system block diagram, and the caps on the TIAGo gripper.
From Figure 4: The six test objects, the TIAGo robot grasping the plastic bottle, and the matching sensor response.
From Figure 4: The six test objects, the TIAGo robot grasping the plastic bottle, and the matching sensor response.
From Figure 6: The three recognition pipelines: end-to-end on raw signals, feature-based with an SVM, and heatmaps wi...
From Figure 6: The three recognition pipelines: end-to-end on raw signals, feature-based with an SVM, and heatmaps with a 2D-CNN.
From Figure 9: Precision, recall and F1 of each approach with 16, 24 and 32 sensors, using 50-sample windows.
From Figure 9: Precision, recall and F1 of each approach with 16, 24 and 32 sensors, using 50-sample windows.
Abstract

Tactile sensing provides information about object properties that can improve robotic grasping and manipulation. However, accurate object recognition during the early stages of grasping remains challenging, particularly under strict real-time constraints. This study presents a flexible, distributed piezoelectric tactile sensing system for early-stage object recognition. The system was experimentally evaluated using six everyday objects and three recognition approaches: feature-based classification using support vector machines (SVMs), raw-signal classification using 1D convolutional neural networks (1D-CNNs) and convolutional recurrent neural networks (C-RNNs), and spatiotemporal heatmap classification using a 2D-CNN. The effects of sensor count and temporal window size on recognition accuracy and real-time feasibility were systematically investigated. Increasing the number of sensing elements improved recognition performance across all approaches. The best-performing method, based on spatiotemporal heatmaps and a 2D-CNN, achieved F1-scores of 0.9796, 0.9846, and 0.9903 using 16, 24, and 32 sensors, respectively. The optimal configuration, comprising 32 sensors and a 100-ms temporal window (50 samples), was deployed on an embedded edge platform, achieving an end-to-end recognition latency of approximately 101 ms from contact onset and an energy consumption of 63.8 microjoules per inference. These results demonstrate the potential of distributed piezoelectric tactile sensing for accurate, energy-efficient, and real-time object recognition during robotic grasping.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.