Machine-learned Adaptive Switching in Voluntary Lower-limb Exoskeleton Control: Preliminary Results

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Lower-limb exoskeletons often use fixed control strategies, limiting adaptability to user intent.
• Switching between exoskeleton walking modes can increase user effort and cognitive load.

Purpose of the Study:

• Investigate temporal-difference learning and general value functions for predicting user walking mode selection in exoskeletons.
• Reduce effort and cognitive load during mode transitions in lower-limb exoskeletons.
• Develop an adaptive control strategy for real-time exoskeleton applications.

Main Methods:

• Employed temporal-difference learning and general value functions to predict user's next walking mode.
• Utilized device-centric and room-centric measurements to learn user switching preferences.
• Tested with a user controlling the Indego exoskeleton across five distinct walking modes.

Main Results:

• The adaptive strategy, using a machine-learned switching list, decreased required manual switches by an average of 42.44% compared to static lists.
• The method is computationally inexpensive and capable of real-time, temporally extended predictions.
• Successfully learned and predicted user switching preferences based on movement similarities.

Conclusions:

• Temporal-difference learning offers a promising, computationally inexpensive approach for adaptive exoskeleton control.
• This technique can significantly reduce manual mode switches, improving user experience and reducing cognitive load.
• The findings pave the way for real-time adaptive control in lower-limb exoskeleton applications.