This talk presents two recent efforts where machine learning addresses fundamental challenges in wearable robot modeling and control. The first contribution addresses the robot side: a hybrid Neural ODE framework for modeling artificial muscle dynamics that embeds physical structure into a learned model, enabling reliable stiffness control over a 126–176 N/mm range. The second addresses the human side: a transformer-based neural decoder that estimates joint impedance directly from EMG, learning time-varying stiffness and equilibrium position under a least-action prior — the assumption that voluntary movement is energetically efficient. These two contributions reflect a common principle: that both sides of the human-robot interface — the robot's dynamics and the user's intent — must be learned rather than prescribed. The talk closes with an emerging direction: Hebbian self-organization of spinal-like reflex networks, which may ultimately replace both learned models with a single continuously adapting system, removing the need for supervised training altogether.
Jonathan Realmuto is an assistant professor in the department of Mechanical Engineering at the University of California, Riverside and a visiting scientist at Children's Hospital Orange County. Together with his research group, the Bionic Systems Laboratory, he designs, builds, and experimentally tests wearable robots and collaborative robots.