Researchers from the Ningbo Institute of Materials Technology and Engineering of the Chinese Academy of Sciences (CAS), in collaboration with the Hong Kong Polytechnic University, have developed a robust feature selection method to improve interaction stiffness estimation for human-robot collaboration.

Their study was published in IEEE Transactions on Industrial Electronics on June 23.

Embodied intelligence and humanoid robots are increasingly used in real-world tasks. In addition to learning human motion trajectories, robots must also master contact skills such as compliance, force adaptation, and stiffness regulation.

In contact-intensive tasks such as assembly and wiping, interaction stiffness determines how softly or firmly a robot should interact with its environment. This is crucial for learning from human demonstrations. However, stiffness estimation typically relies on multimodal sensing systems, and surface electromyography (sEMG) signals are often compromised by muscle crosstalk, motion artifacts, and other noise.

To address this issue, the researchers proposed a noise-free maximum-relevance and minimum-redundancy (MRMR) method driven by extreme value theory (EVT). This method uses EVT to estimate the noise censoring threshold without requiring a predefined confidence level and introduces a noise-free metric to evaluate the similarity of noisy features.

By maximizing noise-free relevance while minimizing noise-free redundancy, NF-MRMR selects compact and informative feature subsets from high-dimensional data affected by unknown noise.

The researchers validated NF-MRMR on 15 benchmark datasets from manufacturing, pharmacy, image recognition, and other fields. "It outperformed 11 representative feature selection methods, achieving the highest average classification accuracy across all classifiers," said Prof. CHEN Silu, corresponding author of the study.

The researchers also applied NF-MRMR to a human-robot collaborative wiping task. Using only 10 selected sEMG features, NF-MRMR reconstructed continuous interaction stiffness, reducing the mean absolute error by about 37.73% compared with three baseline methods. They then used the estimated stiffness to guide a robot's autonomous wiping of traces with different pressure levels.

This study provides a data-driven tool for extracting reliable human-robot interaction cues from noisy physiological signals. It may support stiffness-aware skill learning for humanoid robots and other embodied robotic systems performing contact-rich tasks such as polishing, assembly, surface finishing, and human-guided skill transfer.

This work was supported by the National Key R&D Program of China, the National Natural Science Foundation of China, the Zhejiang Key R&D Program, and others.