Structure of the Ankle Exoskeleton Learning Algorithm
The primary goal of this project was to develop an intelligent torque control system for an ankle exoskeleton aimed at simultaneously minimizing user metabolic cost and maximizing user-perceived comfort and natural movement. This objective addresses a critical challenge in wearable robotics: achieving efficient and personalized assistance that seamlessly integrates with the user's biomechanics, preferences and uses the least energy of user. Effective torque control is crucial for extending battery life in mobile devices, reducing user fatigue, and for adopting assistive robotic technologies. This goal aligns with broader engineering principles of efficiency, optimization, and human-centered design, relevant across various technology domains, including robotics, autonomous systems, and human-machine interfaces.
The original electronics board was bulky, uncomfortable, and chaotic—wires dangled everywhere, and the setup lacked modularity.
I led the redesign of the electronics board, transforming it into a sleek, modular, and wearable system.
The core challenge encountered was significant machine learning overfitting during the development of the torque control model. Initially, the model demonstrated high accuracy on the training dataset, exhibiting a strong ability to predict optimal torque based on input features. However, this performance drastically degraded when evaluated on unseen data, indicating a lack of generalization. This overfitting issue was critical because it would lead to unreliable torque predictions in real-world scenarios, resulting in suboptimal metabolic cost reduction, inconsistent user comfort, or potentially unsafe or unpredictable device behavior. Overfitting stemmed from the complexity of the model relative to the dataset size. Addressing this overfitting was paramount to ensure the model's practical applicability and reliability in diverse user conditions.
My primary contribution was the critical implementation of leave-one-subject-out cross-validation (LOSO CV) to rigorously address the overfitting issue. Recognizing the limitations of traditional train-test splits in this context, particularly the risk of subject-specific biases influencing model evaluation, I strategically chose LOSO CV. This involved partitioning the dataset by subject, iteratively training the model on data from N-1 subjects and validating on the held-out subject, systematically repeating this process for all subjects, and aggregating performance metrics across all folds to obtain a robust estimate of generalization performance. This methodology ensured that the model's performance was evaluated on data completely independent from the training set, providing a far more realistic assessment of its ability to generalize to new users and conditions. The selection of LOSO CV was driven by the need to minimize subject-specific variability and the fact that the dataset we are dealing with are small. If the dataset was larger, LOSO CV would not be a good choice since it takes too much time and energy.
Structure of the Ankle Exoskeleton Learning Algorithm
Results: The implementation of leave-one-subject-out cross-validation demonstrably mitigated the overfitting issue. Validation performance, as measured by Mean Squared Error, improved by 28% during LOSO CV compared to previous validation methods. The gap between training and validation performance was significantly reduced, indicating improved generalization.
The model now exhibits a more consistent and reliable performance across different subjects during cross-validation, suggesting a significantly enhanced ability to generalize to new users and conditions.
Next Steps: While overfitting was effectively addressed, future work focuses on enhancing the model's robustness to potential outlier offline datasets. This includes: 🚀Outlier Detection and Preprocessing: Implementing systematic methods for identifying and handling outlier data points within the offline datasets. This may involve statistical outlier detection methods (e.g., Z-score, IQR-based methods), visual inspection of data distributions, or domain-specific outlier detection techniques related to biomechanics or sensor data. 🚀Real-World Validation: Crucially, transitioning to real-world validation with online data collection and testing to assess model performance in dynamic and less controlled environments, which may uncover new types of outliers or challenges not present in offline datasets.
Score Improvement Data created by MATLAB