Transformer-Based Robust Tactile Object Recognition under Sensor Faults through Training, Adaptation, and Correction

Tactile sensors embedded in robotic systems are vulnerable to hardware-level faults such as short-circuits (saturated high-value output, e.g., 5 N) and open-circuits (zero- value output), which may arise from realistic physical failures such as broken signal lines or internal disconnections. In this work, I explicitly model such hardware-induced faults based on the physical wiring structure of tactile sensor arrays and investigate their impact on object recognition accuracy. I first perform a sensitivity analysis to identify sensor regions where abnormal values (e.g., 0 N or 5 N) have particularly strong influence on recognition performance. Building on these insights, I design a robust recognition framework using a Transformer- based recognition model, and evaluate three strategies: (1) fault- aware training with simulated failure patterns, (2) selective removal of known faulty sensor data, and (3) online correction of abnormal values by converting them into statistically neutral values with minimal impact on the learned force distribution. The experiments show that while fault injection during training provides some robustness, it is difficult to generalize across all possible failure patterns. In contrast, fault detection and correction at inference time significantly restore recognition accuracy under severe fault conditions. These results highlight the necessity of integrating both hardware fault modeling and adaptive inference mechanisms for reliable tactile perception in real-world robotic systems.