Optimizing Locomotor Task Sets in Biological Joint Moment Estimation for Hip Exoskeleton Applications
Jimin An, Changseob Song, Eni Halilaj, Inseung Kang · 2024
Read Optimizing Locomotor Task Sets in Biological Joint Moment Estimation for Hip Exoskeleton Applications by Jimin An, Changseob Song, Eni Halilaj, Inseung Kang (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-07823.txt.
Introduction
Jimin An and co-authors shrink the data burden for hip exoskeletons in a 2024 paper. Estimating a wearer's hip joint moment from sensors takes deep learning. Such models usually demand long lab sessions across many tasks. Their strategy clusters movement tasks by biomechanical features. It then keeps only a minimal representative set for training.
A network trained on the slim set matched full-set accuracy across subjects. Its error stayed near zero point three newton-meters per kilogram. It also clearly beat training on cyclic tasks alone. Readers who build wearable robots get a way to collect far less data. They keep control quality while cutting cost and time.
Catalog record
| CREATOR | Jimin An, Changseob Song, Eni Halilaj, Inseung Kang |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 25 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-07823
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Optimizing Locomotor Task Sets in Biological Joint Moment Estimation for Hip Exoskeleton Applications Jimin An, Changseob Song, Eni Halilaj, Inseung Kang Abstract Accurate estimation of a user’s biological joint moment from wearable sensor data is vital for improving exoskeleton control during real-world locomotor tasks. However, most state-of-the-art methods rely on deep learning techniques that necessitate extensive in-lab data collection, posing challenges in acquiring sufficient data to develop robust models. To address this challenge, we introduce a locomotor task set optimization strategy designed to identify a minimal, yet representative, set of tasks that preserves model performance while significantly reducing the data collection burden. In this optimization, we performed a cluster analysis on the dimensionally reduced biomechanical features of various cyclic and non-cyclic tasks. We identified the minimal viable clusters (i.e., tasks) to train a neural network for estimating hip joint moments and evaluated its performance. Our cross-validation analysis across subjects showed that the optimized task set-based model achieved a root mean squared error of 0.30±0.05 Nm/kg. This performance was significantly better than using only cyclic tasks (p<0.05) and was comparable to using the full set of tasks. Our results demonstrate the ability to maintain model accuracy while significantly reducing the cost associated with data collection and model training. This highlights the potential for future exoskeleton designers to leverage this strategy to minimize the data requirements for deep learning-based models in wearable robot control. I Introduction Lower-limb exoskeletons have been shown to effectively enhance human mobility, thereby improving quality of life [ 1 ] . These robotic systems can augment users and enhance gait performance, for example, by reducing energy expenditure in healthy individuals [ 2 ] and improving spatiotemporal gait metrics in patient populations [ 3 ] . Among different lower-limb exoskeletons, hip exoskeletons present a particularly compelling value proposition due to the critical role of the hip in generating substantial mechanical power during locomotion. Unlike the ankle joint, which benefits from an efficient muscle-tendon mechanism, the hip is mechanically less efficient, requiring greater muscle work (i.e., energy) to complete locomotor tasks [ 4 ] . Control strategies for robotic exoskeletons have significantly advanced over the past decade. Initially relying on heuristic, rule-based controllers utilizing onboard wearable sensors [ 5 , 6 ] , recent approaches now directly leverage the user’s physiological state, such as biological joint moments, to control the exoskeleton more effectively [ 7 ] . Biological torque controllers directly estimate the user’s biological joint moments and apply a reshaped, scaled envelope as the assistance profile [ 8 , 9 ] . This method leverages deep learning techniques with wearable sensing as model input and has demonstrated significant performance benefits [ 10 , 11 ] . The range of activities that can be assisted by exoskeletons has expanded significantly, progressing from simple continuous walking [ 12 , 13 ] to dynamic transitions between locomotion modes [ 14 , 11 ] , and now including various non-cyclic movements [ 15 , 10 ] . These advancements hold the promise of enabling more natural hip exoskeleton assistance during daily locomotor tasks. However, a significant bottleneck of deep learning approaches is their data-intensive nature, requiring large datasets that comprehensively represent various activity modes, along with ground-truth joint moments obtained through motion capture. Studies have demonstrated that user-independent models typically require data from at least a dozen participants completing multiple tasks [ 16 , 12 ] , with some relying on even larger groups and more diverse tasks [ 11 , 10 ] . While comprehensive datasets are essential for ensuring reliable model performance across various activities, the associated costs and resources are substantial. Furthermore, for patient populations, such data collection can be even more challenging due to physical limitations, time constraints, and logistical barriers. These challenges highlight the necessity for an automated solution that reduces the burden of data collection while not compromising model performance. Fig. 1: Task Selection and Model Training. Average joint moment, angle, and velocity profiles were used for principal component analysis (PCA), k-means clustering, and task-weight analysis to determine a smaller optimized set of tasks that were representative of the full set. The time-series data for each training dataset condition (all tasks, optimized tasks, and cyclic tasks) were used for the fully connected neural network (FCNN) model training with leave-one-subject-out cross-fold validation. Few studies have verified that optimal model performance can be achieved with data from a reduced set of tasks [ 10 , 15 ] . These studies employed an iterative forward task selection approach, incrementally adding tasks to the training dataset until further additions no longer improved model performance. While these results suggest that a full dataset is not always necessary during training, they do not clarify which tasks are most relevant and useful or explain why certain tasks are more effective in developing user-independent joint moment estimators. Without a deeper analysis beyond model performance, these approaches provide limited insight into the relationships between tasks and their biomechanical relevance. This gap emphasizes the need for a systematic approach to identify a minimal task set that ensures reliable model training and performance while reducing data collection requirements. Given the similarities among human movement profiles across various locomotor tasks, collecting data with finely discretized task separation may not always be necessary. Instead, grouping tasks based on shared biomechanical features might be a better strategy for improving efficiency and applicability in various contexts. Prior studies have applied dimensionality reduction and clustering methods to biomechanical data analysis for applications such as locomotion phase classification [ 17 , 18 ] , kinematic pattern recognition [ 19 ] , gait differentiation [ 20 ] , and movement analysis [ 21 , 22 , 23 ] . Building on prior literature, grouping tasks using a clustering approach based on shared biomechanical features offers a systematic solution for identifying redundant tasks that provide similar information. By selecting representative tasks from each cluster and training the model on these tasks, we can significantly reduce data collection efforts while preserving essential biomechanical patterns needed for accurate joint moment estimation. In this work, we propose a strategy for optimizing locomotor task selection through clustering analysis. We hypothesized that a model trained on an optimized task set would achieve hip joint moment estimation comparable to a model trained on the full dataset. Our underlying rationale is that the latent space representation of these tasks likely exhibits significant overlap for certain tasks due to their shared biomechanical features. To test this hypothesis, we trained a neural network using three different datasets to evaluate the effectiveness of our task set reduction: 1) the full set of tasks, 2) an optimized reduced set of tasks, and 3) only cyclic tasks. The model estimated hip joint moments using sensor data commonly available for hip exoskeletons, such as encoders and inertial measurement units. Our study results contribute to the field by demonstrating that a biological torque controller for hip exoskeletons can be developed using data from fewer tasks, while also identifying key tasks for future data collection efforts. II Methods II-A Dataset and Data Pre-Processing We used an open-source biomechanics dataset of 12 healthy young individuals (7 males, 5 females, age of 21.8 \pm 3.2 years, height of 176.7 \pm 8.6 cm, and