SpikeTouch: Optimizing Spike Neural Networks for Tactile Perception

Tactile perception enables systems to sense and interpret physical properties such as shape, material, pressure, and texture. It is a key capability for emerging applications like robotic surgery and assistive robotics. Existing solutions to tactile perception typically rely on computation-intensive deep neural networks, which require high-performance computing resources unavailable on embedded and battery-powered devices. Spiking neural networks (SNNs) offer a promising, energyefficient alternative, but their practical adoption remains limited due to the lack of efficient and deployable neuromorphic solutions. We present SpikeTouch, a software framework designed to reduce the computational overhead of SNNs for tactile perception on neuromorphic hardware. SpikeTouch offers three key optimizations tailed to SNN-based tactile perception:

(1) a spike encoding scheme that balances precision and computational cost; (2) a systematic method for extracting multidimensional tactile features; and (3) a training strategy that minimizes quantization errors to improve performance and reduce memory usage. We evaluate SpikeTouch on the Tianjic neuromorphic chip using representative tactile perception workloads. Our results show that SpikeTouch achieves high recognition accuracy for 30 objects and their material stiffness , with an average accuracy of 92.92% and 92.35%, respectively. It is also highly energy and computationally efficient, operating at just 1 Watt of power -significantly lower than the hundreds or thousands of Watts typically required by a GPU -and delivering a response in under 20 ms -well below the average human reflex time of over 100 ms.

CCS Concepts: • Human-centered computing → Ubiquitous and mobile computing.