Motion Intention Prediction for Lumbar Exoskeletons Based on Attention-Enhanced sEMG Inference.

Area of Science:

• Robotics
• Artificial Intelligence
• Biomechanics

Background:

• Exoskeleton robots augment human capabilities via mechanical coupling and assistive torques.
• Effective human-robot cooperative control requires accurate resolution of human kinematic intent.

Purpose of the Study:

• To introduce a lumbar spine-assisted exoskeleton design utilizing Variable-Stiffness Pneumatic Artificial Muscles (VSPAM).
• To develop a dynamic adaptation mechanism for seamless human-robot cooperative control.
• To enhance kinematic intent resolution through multimodal fusion and attention mechanisms.

Main Methods:

• Designed a lumbar exoskeleton with VSPAM for assistive torques.
• Developed a dynamic adaptation mechanism for pneumatic drive and human intent.
• Proposed a multimodal fusion architecture integrating VGG16 and LSTM networks with self-attention and cross-attention mechanisms for intent resolution.

Main Results:

• Achieved 96.1% ± 1.2% motion classification accuracy in experimental validation.
• Demonstrated effective fusion of visual and kinematic features using cross-attention.
• Showcased the ability to capture global spatio-temporal dependencies with self-attention mechanisms.

Conclusions:

• The developed exoskeleton and intent resolution system offer a novel technical solution for rehabilitation robotics and industrial assistance.
• The multimodal fusion architecture effectively mitigates unimodal perception limitations.
• The self-attention mechanism enhances the capture of complex human movement patterns.