Bio-Inspired Internal Representations of Tactile Sensation, Pain, and Damage for Artificial Skin Using Spatio-Temporal Anomaly Detection.

In recent years, the deployment of robots in human-centric environments has necessitated the development of artificial skins that integrate safety and durability. Traditional damage detection often relies on raw signal thresholds, lacking the functional integration of touch, pain, and damage found in biological systems. This study proposes a bio-inspired artificial skin model that separately evaluates these three states through a spatio-temporal anomaly detection framework. We developed an unsupervised model combining a Convolutional Autoencoder (CAE) and Convolutional LSTM (ConvLSTM) to learn the latent representations of tactile maps from intact skin. By quantifying spatial reconstruction and temporal prediction errors, the system generates individual scores for touch, pain, and damage. Pain is defined as an abstract signal of instantaneous abnormality, while damage is identified as a persistent structural deviation. We implemented a dynamic thresholding mechanism mimicking biological sensitization and recovery, with damage detection gated by a pain-flag constraint to minimize false positives. Experimental results across various conditions-including incisions (3-6 cm) and abrasions (10-30 times)-demonstrate that the model can distinguish between momentary noxious stimuli and sustained structural degradation. Quantitative evaluation shows that the proposed model achieves an Area Under the Curve (AUC) of 0.653, outperforming a threshold-based baseline and maintaining zero false positives under strong, non-damaging contact. Specifically, the system successfully mimics biological aftereffects and the pain-gating mechanism, where damage is only assessed in the presence of a pain-related trigger. This research provides a scalable, software-driven foundation for robot self-protection that overcomes the implementation constraints of hardware-dependent neuromorphic systems.