[Paper Review] Compressed Learning for Tactile Object Classification

Brayden Hollis, Stacy Patterson+1 more|arXiv (Cornell University)|Sep 24, 2016

Advanced Sensor and Energy Harvesting Materials15 references3 citations

TL;DR

This paper proposes compressed learning for tactile object classification using compressed sensing to reduce data acquisition, wiring complexity, and processing time while maintaining high classification accuracy. By classifying directly from compressed tactile signals—without full reconstruction—it achieves near-100% accuracy even at 64× compression and with as little as 1% training data.

ABSTRACT

The potential of large tactile arrays to improve robot perception for safe operation in human-dominated environments and of high-resolution tactile arrays to enable human-level dexterous manipulation is well accepted. However, the increase in the number of tactile sensing elements introduces challenges including wiring complexity, power consumption, and data processing. To help address these challenges, we previously developed a tactile sensing technique based compressed sensing that reduces hardware complexity and data transmission, while allowing accurate reconstruction of the full-resolution signal. In this paper, we apply tactile compressed sensing to the problem of object classification. Specifically, we perform object classification on the compressed tactile data. We evaluate our method using BubbleTouch, our tactile array simulator. Our results show our approach achieves high classification accuracy, even with compression factors up to 64.

Motivation & Objective

• Address the challenges of high wiring complexity, data transmission, and processing overhead in large-scale tactile sensor arrays.
• Enable efficient tactile object classification by bypassing full signal reconstruction and classifying directly from compressed measurements.
• Demonstrate that compressed sensing preserves classification performance even at high compression ratios and with limited training data.
• Explore the feasibility of using compressed signals for downstream tasks like grasping, where full signal recovery remains possible.

Proposed method

• Apply compressed sensing to tactile data acquisition, sampling and compressing tactile signals simultaneously using random measurement matrices.
• Use a soft-margin support vector machine (SVM) to classify objects directly from the compressed signals, avoiding reconstruction.
• Train and test the classifier on compressed signals derived from full-resolution tactile snapshots of objects pressed onto a square taxel array.
• Simulate tactile data using BubbleTouch, a tactile array simulator, to generate realistic tactile signals under controlled conditions.
• Evaluate performance across varying compression factors (up to 64×) and training set sizes (from 1% to 100%).
• Compare classification accuracy of compressed signals against full-resolution raw signals and low-resolution raw signals of equivalent size.

Experimental results

Research questions

• RQ1 Can tactile object classification be accurately performed directly on compressed tactile signals without full reconstruction?
• RQ2 How does classification accuracy vary with increasing compression ratios (e.g., 64×) using compressed learning?
• RQ3 How much training data is required for high-accuracy classification when using compressed signals versus raw low-resolution signals?
• RQ4 Does compressed learning outperform raw low-resolution signals in terms of classification accuracy under equivalent signal dimensions?

Key findings

• The proposed compressed learning approach achieves near-100% classification accuracy even at 64× compression, demonstrating robustness to high data reduction.
• With only 1% of the training data, compressed signals of size 16 maintain over 70% accuracy, significantly outperforming raw signals of the same size (under 60%).
• At 3% training data, compressed signals of size 64 achieve over 80% accuracy, showing strong generalization despite limited data.
• The confusion matrix reveals the highest misclassification occurs between the volleyball and basketball (25% confusion), due to similar spherical shapes and contact radii.
• The method maintains performance comparable to full-resolution signals, with deviations increasing only slightly as training data diminishes, consistent with compressed learning theory.
• Compressed signals outperform raw signals of equivalent size, indicating that compression not only reduces data volume but may also suppress noise, improving classification robustness.