Researchers at the Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences, have developed a novel robust feature selection framework that significantly enhances interaction stiffness estimation for human-robot collaboration. Published recently in IEEE Transactions on Industrial Electronics, the study addresses a critical bottleneck in embodied AI and humanoid robotics: the accurate perception of contact dynamics during physical tasks such as polishing, assembly, and surface wiping. Traditional approaches to estimating interaction stiffness rely heavily on multi-modal sensor systems, including motion capture and surface electromyography. However, surface electromyography signals are inherently vulnerable to muscle crosstalk, motion artifacts, and unpredictable environmental noise, which destabilizes feature extraction and degrades control precision. To overcome these limitations, the research team engineered an Extreme Value Theory-driven Noise-Free Maximum Relevance Minimum Redundancy algorithm. By leveraging Extreme Value Theory to dynamically estimate noise truncation thresholds without preset confidence levels, the method constructs a noise-free similarity metric to evaluate feature redundancy. It subsequently optimizes feature selection by maximizing relevance while minimizing redundancy, effectively isolating compact, high-information subsets from highly contaminated physiological data streams. Validation across fifteen benchmark datasets confirmed the algorithm’s statistical robustness. In practical human-robot collaborative wiping experiments, the system successfully reconstructed continuous interaction stiffness using merely ten filtered signals. This streamlined approach reduced the mean absolute error by approximately 37.73 percent compared to three established baseline techniques, enabling the robotic system to autonomously modulate its compliance and adapt to varying pressure traces in real time. The methodology establishes a reliable, data-driven pipeline for extracting stable human-robot interaction cues from noisy biological signals. By decoupling critical tactile information from signal degradation, the framework provides a scalable foundation for stiffness-aware skill learning and imitation. Its deployment is expected to accelerate the transition of humanoid and embodied systems into complex industrial environments where precise force modulation and physical contact are paramount. The initiative received financial backing from the National Key Research and Development Program of China and the National Natural Science Foundation of China.