Title: Continual Imitation Learning for Prosthetic Limbs

Abstract: Lower limb amputations and neuromuscular impairments severely restrict mobility, necessitating advancements beyond conventional prosthetics. Motorized bionic limbs offer promise, but their utility depends on mimicking the evolving synergy of human movement in various settings. In this context, we present a novel model for bionic prostheses' application that leverages camera-based motion capture and wearable sensor data, to learn the synergistic coupling of the lower limbs during human locomotion, empowering it to infer the kinematic behavior of a missing lower limb across varied tasks, such as climbing inclines and stairs. We propose a model that can multitask, adapt continually, anticipate movements, and refine. The core of our method lies in an approach which we call -- multitask prospective rehearsal -- that anticipates and synthesizes future movements based on the previous prediction and employs a corrective mechanism for subsequent predictions. We design an evolving architecture that merges lightweight, task-specific modules on a shared backbone, ensuring both specificity and scalability. We empirically validate our model against various baselines using real-world human gait datasets, including experiments with transtibial amputees, which encompass a broad spectrum of locomotion tasks. The results show that our approach consistently outperforms baseline models, particularly under scenarios affected by distributional shifts, adversarial perturbations, and noise.