Investigating the Generalizability of Assistive Robots Models over Various Tasks

TL;DR

This work tackles the problem of generalizability in data-driven dynamic models for assistive exoskeletons by learning the interaction dynamics with six regression methods using inputs $\tilde{x}=[x,u,v]$ to predict $\Delta x$ under Gaussian noise, i.e., $x_{t+1}=x_t+f(x_t,u_t,v_t)+\eta_t$ with $\Delta x_t=x_{t+1}-x_t$. Six tasks (Horizontal, Vertical, LR, RL, Eating, Pushing) and data from three subjects are used to evaluate cross-task transfer via five-fold cross-validation. The results show XGBoost provides the strongest cross-task generalizability (about $84.93\%$ on average), with Gaussian Process Regression close behind (about $82.31\%$), and the horizontal task (H) yielding the best cross-task generalization across models; MLP trails others. These findings offer practical guidance for selecting task sets and regression methods to build more data-efficient, transferable assistive-control systems, which is crucial for real-world use by individuals with disabilities. $x_{t+1}=x_t+f(x_t,u_t,v_t)+\eta_t$, $\Delta x_t=x_{t+1}-x_t$, and cross-task generalizability thresholds (e.g., $80\%$ R^2) provide concrete benchmarks for future work.

Abstract

In the domain of assistive robotics, the significance of effective modeling is well acknowledged. Prior research has primarily focused on enhancing model accuracy or involved the collection of extensive, often impractical amounts of data. While improving individual model accuracy is beneficial, it necessitates constant remodeling for each new task and user interaction. In this paper, we investigate the generalizability of different modeling methods. We focus on constructing the dynamic model of an assistive exoskeleton using six data-driven regression algorithms. Six tasks are considered in our experiments, including horizontal, vertical, diagonal from left leg to the right eye and the opposite, as well as eating and pushing. We constructed thirty-six unique models applying different regression methods to data gathered from each task. Each trained model's performance was evaluated in a cross-validation scenario, utilizing five folds for each dataset. These trained models are then tested on the other tasks that the model is not trained with. Finally the models in our study are assessed in terms of generalizability. Results show the superior generalizability of the task model performed along the horizontal plane, and decision tree based algorithms.

Figures (6)

• Figure 1: Myopro prosthetic augmented with IMU sensors for data collection. Figure annotation highlights the placement of sEMG sensors and the rotational axes of APDM Opal IMU sensors.
• Figure 2: Overview of the diverse tasks employed for data acquisition from test subjects, including Horizontal (H), Vertical (V), diagonal from Left leg to Right eye (LR), diagonal from Right leg to Left eye (RL), Eating (E), and Pushing (P).
• Figure 3: Mean and standard deviation of the first two input features (elbow angle and angular velocity) calculated with $99\%$ confidence interval from four trials for each of the six tasks, with data collected from three test subjects distinguished by different colors from left to the right.
• Figure 4: Performance evaluation of each model, where training is conducted on the home node and testing is performed on the destination node data. Edge color intensity inversely correlates with the models' ability to generalize; a brighter edge shows lower generalizability.
• Figure 5: The generalizability of different task data sets within each algorithm, using a R-squared value of 80% as the threshold for acceptable performance.