MAD-CNN: High-Sensitivity and Robust Collision Detection for Robots with Variable Stiffness Actuation

Abstract Safe and efficient human-robot collaboration (HRC) relies on proactive collision avoidance and robust detection of unexpected contacts. Although existing learning-based approaches excel in specific areas, they often encounter challenges with robots equipped with variable stiffness actuators—a crucial aspect for ensuring safe and adaptable HRC. To address this, we present MAD-CNN, a novel Modularized Attention-Dilated Convolutional Neural Network designed for highly sensitive and robust collision detection in these adaptable robots. MAD-CNN leverages a dual inductive bias mechanism, combining modularization and dilated convolution, to significantly reduce data requirements. This capability facilitates efficient training even with limited collision datasets, offering a substantial advantage for real-world implementation. Additionally, an attention module prioritizes relevant features, enhancing the network's robustness against variations in stiffness levels—a core challenge in variable stiffness robots.This powerful combination unlocks exceptional performance for MAD-CNN. In extensive evaluations, it achieved zero missed collisions across 516 events, with a minimal detection delay averaging just 12.05 milliseconds and 20\% fewer false positives compared to baseline methods. Remarkably, this outstanding performance was attained even when trained on a four-minute dataset focused on the highest stiffness setting, underscoring MAD-CNN's data efficiency and adaptability across diverse stiffness configurations. By enabling accurate and sensitive collision detection across the spectrum of variable stiffness conditions, MAD-CNN paves the way for safer and more reliable HRC in various fields, including manufacturing, healthcare, and assistive robotics.