A Weight-Aware-Based Multisource Unsupervised Domain Adaptation Method for Human Motion Intention Recognition.

Area of Science:

• Robotics
• Machine Learning
• Human-Computer Interaction

Background:

• Accurate human motion intention (HMI) recognition is crucial for natural human-robot interaction in exoskeleton systems.
• Individual motor characteristics cause performance degradation in classifiers trained on different subjects (domain shift).
• Unsupervised domain adaptation (UDA) addresses this but current methods struggle with multiple, diverse source subjects.

Purpose of the Study:

• To develop a novel UDA theory and algorithm for HMI recognition that accounts for differences among multiple source subjects.
• To improve the accuracy and generalization ability of HMI recognition systems in the presence of domain shift.

Main Methods:

• Extended margin disparity discrepancy (MDD) to multisource UDA theory.
• Proposed a weight-aware-based multisource UDA algorithm (WMDD) incorporating adaptive source domain weights based on MDD.
• Employed a lightweight network for real-time classification and adversarial learning for enhanced generalization.

Main Results:

• The developed multisource UDA theory guarantees generalization error on the target subject.
• The WMDD algorithm effectively bridges theory and practice by transforming theory into an optimization problem.
• Extensive experiments demonstrate WMDD's superior performance over existing UDA methods in HMI recognition.

Conclusions:

• The proposed WMDD algorithm offers a robust solution for HMI recognition by effectively handling multisource domain shifts.
• The theoretical framework provides a guarantee for generalization performance.
• WMDD enhances the comfort and naturalness of human-robot interaction in exoskeleton applications.