Object recognition is crucial for survival in many species, including humans. Current AI-based object recognition systems heavily rely on visual input, making them vulnerable in conditions with poor lighting or obstructions. This study explores an alternative approach using tactile and olfactory sensing, drawing inspiration from the star-nosed mole, a mammal exceptionally adept at object recognition in dark, underground environments. The star-nosed mole's sophisticated tactile and olfactory systems are seamlessly integrated, allowing for rapid and accurate object identification with minimal reliance on vision. This bio-inspired approach offers several potential advantages: compact sensory components, high accuracy, robustness to environmental interference, high efficiency, and low power consumption. This paper details the design, fabrication, and testing of a bionic tactile-olfactory sensing system designed to replicate this biological success, focusing on the potential application of human identification during rescue missions in challenging environments.

Existing object recognition methods predominantly utilize visual information processed by deep convolutional neural networks (CNNs). While highly successful in many contexts, these methods struggle in non-visual environments or scenarios with obstructions. Research into multi-modal sensing, incorporating data from other sensory modalities like somatosensory and auditory inputs, shows promise. However, the combined use of tactile and olfactory data for object recognition is less explored due to the challenges in integrating data from these diverse sensor types. The unique sensory capabilities of the star-nosed mole, relying heavily on tactile and olfactory information, highlight the potential of this combined approach for robust object recognition in challenging environments. Previous research has focused on understanding the mole's sensory organs and neural processing, providing insights for the design of artificial tactile and olfactory sensors. This study builds upon this prior work by creating a comprehensive system that integrates these sensors and implements a bio-inspired machine learning algorithm for accurate object recognition.

The researchers developed a tactile-olfactory sensing array mimicking the star-nosed mole's unique nose structure. The system incorporates a mechanical hand with 70 force sensors distributed across five fingertip arrays and an olfactory array with six gas sensors sensitive to different gases. The sensors were fabricated using silicon-based microelectromechanical systems (MEMS) technology to achieve high sensitivity and stability. The design of the force sensors ensures high resolution of tactile information, while the gas sensors provide customizable odor perception, adapting to the environment. The sensors' responses were characterized for sensitivity, stability, and response time under various conditions. A bio-inspired olfactory-tactile (BOT) associated machine-learning algorithm, modeled on the star-nosed mole's neural system, processes the combined tactile and olfactory data. The algorithm consists of three neural network layers: early tactile and olfactory processing, feature extraction with preliminary weight determination for each modality, and multi-sensory fusion. A dataset of 55,000 samples from 11 object types, categorized into five groups (human, olfactory interference, tactile interference, soft, and rigid objects), was created for training and testing the BOT algorithm. To improve performance, the researchers optimized the algorithm by altering eigenvalue extraction methods and data output modes, resulting in several variants: BOT-R, BOT-F, and BOT-M. The performance of these variations were compared with unisensory approaches (tactile-only and olfactory-only recognition). The system's robustness to environmental interference was tested by introducing Gaussian noise into the tactile and olfactory data.

The fabricated force sensors demonstrated high sensitivity (0.375 mV kPa⁻¹), accurately reflecting multiple contact statuses during object interaction. The gas sensors showed rapid response times and high stability over time and temperature changes. The BOT algorithm, particularly the optimized BOT-M variant, achieved a 96.9% accuracy in classifying the 11 object types. Compared to unisensory approaches, BOT-M significantly outperformed both tactile-only (81.9%) and olfactory-only (66.7%) recognition strategies. The system showed excellent tolerance to environmental interference, maintaining high accuracy even with significant Gaussian noise introduced into the sensory data. The system's effectiveness in human identification was validated in a simulated rescue scenario at a fire department test site. The tests involved scenarios with gas interferences (acetone and ammonia), partially buried objects, and damaged sensors. The BOT-M algorithm maintained high recognition accuracy (above 80%) even in these challenging conditions, demonstrating its robustness.

The results demonstrate the feasibility and effectiveness of combining tactile and olfactory sensing for robust object recognition in non-visual environments. The superior performance of the BOT-M algorithm compared to unisensory approaches highlights the importance of multi-modal sensory integration. The system's resilience to environmental interferences such as gas leaks, partial burial, and sensor damage showcases its potential for real-world applications like search and rescue missions. The bio-inspired approach, modeling the neural processing of the star-nosed mole, offers a novel solution for object recognition challenges that surpass the capabilities of traditional vision-based systems. The compact sensor design, low power consumption, and efficient data processing make this approach particularly well-suited for portable and autonomous robotic systems.

This study successfully developed a bio-inspired tactile-olfactory sensing system capable of robust object recognition in non-visual environments. The optimized BOT-M algorithm achieved high accuracy and demonstrated resilience to various environmental interferences. The findings highlight the potential of this technology for applications in challenging environments, especially search and rescue operations. Future research could explore further algorithm optimization, miniaturization of the sensor array, and integration with advanced robotic platforms for improved manipulation and navigation in complex scenarios.

The current study focused on a limited set of 11 objects. Further testing with a wider range of objects and environmental conditions is necessary to validate the system's generalizability. While the system showed high accuracy in the simulated rescue scenario, real-world testing in diverse and unpredictable environments is crucial to confirm its effectiveness in practical scenarios. The algorithm's dependence on a pre-trained model limits its adaptability to entirely new object types or significant changes in environmental conditions. Future work could explore the development of more adaptable and self-learning algorithms.