Mechanoreceptor inspired multi-class hardness classification approach using cost effective tactile sensor for robotic gripper

<p dir="ltr">Perception is an essential component of robotic systems, enabling interaction with the environment through grasping and tactile exploration. Various sensor modalities, including tactile sensors, cameras, and acoustic sensors, have been integrated into robotic platforms to enhance this capability. Among these, tactile sensing and material classification are particularly important, as they aim to replicate the human ability to assess object properties such as texture and hardness by grasping or touch in robotics. Inspired by human mechanoreceptors, bio-inspired tactile systems have been developed to classify material properties through touch and grasping. These sensors have explored various literature which is often relying on customised sensor arrays and complex architectures to improve robotic grasping and material recognition. This study takes inspiration from human mechanoreceptors and introduces a novel approach using cost-effective, easy-to-install, and readily deployable commercial off-the-shelf (COTS) sensors with similar functionality. Unlike previous research that relies on customised sensor arrays with intricate designs, this work investigates multiple COTS sensors for material hardness classification. A qualitative approach is used, selecting objects based on the Shore hardness taxonomy to evaluate sensor performance. The proposed approach primarily utilizes single-sensor configurations for hardness classification while also evaluating multi-sensor data configurations. Additional sensor data were used to emulate mechanoreceptor spike patterns and incorporate them as extra features to assess their impact on classification accuracy. This also aims to emulate spike patterns using COTS sensors data, analogue-to-spike conversion and raster plot techniques. Inspired by mechanoreceptor architecture, sensors were further arranged in layered and alternative topological configurations to explore and identify the optimal topology for hardness classification. Experimental results indicate that a three-sensor configuration (force, vibration, and potentiometer (F, V, P)) achieved an accuracy of 91.9% in binary classification, while individual sensors did not perform efficiently. In terms of spike patterns, certain time windows exhibited distinct patterns but all of them, also in some cases, spike features improved accuracy from 91.9% to 96% two classes, 81% to 86% three class, 79% to 89% for four classes in case of (F, V, P) data. In case of bio-inspired topologies, the PFV configuration performed optimally in 20% of test scenarios achieving 98.8% in three class than FPV which was 96.5%. In offline and online testing, FPV achieved a higher number of correct predictions than PFV. However, PFV demonstrated lower latency (~25-30 ms) compared to FPV (~35-60 ms), making it more suitable for real-time applications. These findings suggest that a bio-inspired COTS-based approach can be useful for hardness classification as shown by offline testing, whereas real-time or online applications require further optimisation or model tuning.</p>