Automatic Detection of Myocontrol Failures Based upon Situational Context Information

Area of Science:

• Human-Computer Interaction
• Robotics
• Machine Learning

Background:

• Myoelectric control systems for assistive devices suffer from unreliability due to signal instability from factors like fatigue and electrode displacement.
• Current systems require constant updates and significant user feedback to maintain performance.
• A need exists for robust failure detection to ensure dependable control.

Purpose of the Study:

• To develop an automatic failure detection method for myoelectric control systems.
• To enhance control system reliability by identifying when predictions become unreliable and model updates are needed.
• To improve the detection of user intention in assistive device control.

Main Methods:

• Implemented a control system enhanced with generative models trained on human demonstrations of a grasping task in Virtual Reality.
• The system learned reach-to-grasp motions and context-aware grasp prediction (power vs. tridigital grasp).
• Failure detection was achieved by measuring the error between new grasp attempts and model predictions.

Main Results:

• The approach effectively detected input commands that did not align with user intentions by measuring prediction errors.
• Statistical analysis showed highly significant differences (p<0.001) in error distributions between successful and failed grasp attempts.
• The method proved effective using both full wrist pose (position and rotation) and rotation-only information.

Conclusions:

• The proposed automatic failure detection method enhances the reliability of myoelectric control systems.
• Generative models learning from demonstration provide a robust way to detect control failures.
• This approach can significantly improve the performance and dependability of assistive devices.