Deep domain adaptation eliminates costly data required for task-agnostic wearable robotic control.

Area of Science:

• Robotics
• Biomechanics
• Machine Learning

Background:

• Data-driven methods are crucial for wearable robots in rehabilitation and augmentation.
• High-quality, device-specific data are essential but costly and difficult to obtain.
• Data scarcity limits the personalization and performance of current wearable robot systems.

Purpose of the Study:

• To develop a framework that overcomes data scarcity in training wearable robot models.
• To leverage simulated sensor data from biomechanical models as an intermediate domain.
• To enable the translation of easily accessible data into data-limited, device-specific domains.

Main Methods:

• Developed and optimized a deep domain adaptation network.
• Replaced costly, device-specific labeled data with open-source and unlabeled exoskeleton data.
• Trained a hip and knee joint moment estimator using the proposed framework.

Main Results:

• The trained estimator achieved performance comparable to models trained with complete device-specific datasets, with only a slight increase in error (11-44%).
• The proposed network significantly outperformed networks without domain adaptation (outperforming by 36-60%).
• Real-time deployment in a hip/knee exoskeleton demonstrated similar estimator performance and reduced user metabolic cost by 9.5-14.6%.

Conclusions:

• The framework effectively overcomes data scarcity by using simulated data as a bridge.
• It enables the training of real-time deployable, deep learning models with limited or no labeled, device-specific data.
• This approach facilitates advancements in wearable robot applications for rehabilitation and human augmentation.