Motivation: The demand for rehabilitation is steadily increasing due to an aging and more active population. By 2050, the number of people aged over 65 is projected to reach 129.8 million in Europe. Additionally, the rising prevalence of degenerative diseases is placing significant strain on existing rehabilitation resources. Robotic-assisted rehabilitation offers a promising solution to deliver repeatable, high-quality therapy while reducing clinical workload.

Paper accepted to IROS 2026 :https://arxiv.org/abs/2603.14160

In this work, we propose a video-based Learning from Demonstration (LfD) framework that enables therapists to remotely teach multi-modal robotic rehabilitation exercises using RGB-D video. By extracting 6-DoF, body-centric trajectories, the system dynamically scales therapeutic motions to safely fit diverse patient anatomies. To ensure engaging physical human-robot interaction, we developed a decoupled hybrid controller that guides the patient's limb through a spatially compliant virtual tunnel along the exercise trajectory. This architecture directly links the exercise execution speed to the patient's own physical effort, supporting Passive, Active-Assisted, and Active-Resistive therapy modes. Combined with an adaptive, on-the-fly safety mechanism that triggers smooth trajectory reversals upon detecting abnormal forces.