Accelerating Constrained Continual Learning with Dynamic Active Learning: A Study in Adaptive Speed Estimation for Lower-Limb Prostheses

Continual Learning is quickly emerging as a fun-damental technique in almost all technical domains. This study develops its application in robotics, with a specific focus on transfemoral prosthetics, where machine learning models are fine-tuned in real-time to better predict ambulatory speed. This process of model adaptation faces several challenges stemming from the necessity of learning fast enough to keep up with real-time gait, while also ensuring sufficient accuracy and plasticity when encountering changing speeds and modalities. To address these challenges, we introduce Dynamic Active Learning (DAL) and Intermittent DAL (IDAL), novel frameworks which employ uncertainty-based sampling, as potential precursor steps to learning in this adaptation pipeline. Our contributions not only provide a robust guarantee that adaptation will occur within the time constraints posed by gait cycles, but also increase the rate of accuracy convergence by 51 %, IDAL has been shown to attain a 4 % lower post-convergence error rate, and maintain 30 % more reliable post-convergence predictions compared to non-AL based methods of adaptation. In developing this system, we assessed numerous uncertainty metrics, finding that the Query by Committee method performs the best, attaining a Spearman Correlation Coefficient of 0.81 with ground truth error. While showcased through transfemoral prosthetics, our results illustrate the wide reaching potential of our DAL systems across diverse robotics applications.