Researcher: Hossein Kaviani rad, Satoshi Endo

Motivation

Hybrid rehabilitation systems combining robotic exoskeletons with functional electrical stimulation (FES) present a powerful opportunity to restore upper‑limb function in people with severe motor impairments, such as stroke survivors and individuals with spinal cord injury. Robotic assistance offers precise and consistent mechanical support, while FES directly engages the neuromuscular system, activating muscles and promoting neuroplasticity.

The challenge is to achieve balanced actuation sharing between the human, the exoskeleton, and FES. Over‑reliance on the robot can reduce active neuromuscular engagement, limiting recovery potential, while excessive FES can cause discomfort, rapid fatigue, and reduced therapeutic effectiveness. The neuromuscular response to FES is inherently time‑varying, affected by muscle fatigue, recruitment patterns, and non‑linear activation dynamics, and varies between individuals and across therapy sessions.

To address this, we need robust, trustworthy control strategies grounded in models that capture the uncertainty, variability, and noise inherent in human–machine interaction, and that adapt in real time to changes in both the human and device behaviour.

Research questions

• How can we model hybrid human–machine systems in a way that captures variability in mechanical and physiological response, including the time‑varying properties of FES?
• How can control strategies dynamically manage actuation redundancy between voluntary effort, FES, and robotic assistance to maximise functional recovery?
• How can uncertainty in system behaviour be estimated and incorporated into control decisions to maintain safety and trust?
• How can these models adapt over the course of rehabilitation, progressively shifting the balance of effort toward the patient as capability improves?

Approach

Our approach focuses on developing probabilistic, time‑adaptive models of the hybrid exoskeleton–FES system that explicitly account for variability in human motor performance and neuromuscular response. These models continuously monitor changes in FES effectiveness, mechanical output, and voluntary contribution, updating their internal representation in real time.

By incorporating uncertainty quantification into the control loop, the system can recognise when confidence in its predictions is low and adjust its assistance strategy accordingly. This enables it to allocate torque dynamically between the human, FES, and robot, ensuring just enough support to achieve stable, high‑quality movement while keeping the patient actively engaged. In stroke rehabilitation, this allows the controller to adapt as strength and coordination improve, progressively reducing robotic support and encouraging voluntary movement.

Beyond real‑time adaptation, this framework also guides targeted data collection for model refinement, identifying which conditions require further clinical measurement and providing insights that can inform personalised rehabilitation planning.

Key results and achievements

• Defined a redundancy‑aware control paradigm that dynamically balances torque sharing between voluntary effort, FES‑induced activation, and robotic assistance.
• Developed stochastic, uncertainty‑aware modelling concepts that capture the time‑varying behaviour of FES‑driven muscle activation.
• Demonstrated in pilot studies that adaptive actuation sharing can improve movement smoothness and completeness compared to FES‑only or robot‑only approaches.
• Preserved neuromuscular engagement while maintaining stability and comfort, supporting long‑term recovery potential.