Researcher: Satoshi Endo

Motivation

When humans perform movements, their motor output is shaped not only by planned feedforward commands but also by feedback mechanisms that maintain stability and adapt to changing conditions. These feedback components arise from intrinsic viscoelastic properties of muscles, short‑ and long‑latency reflexes, and voluntary corrections. In dynamic or unpredictable environments, rapid intrinsic and reflexive responses are critical for stability, as voluntary feedback alone is too slow.

The mechanical impedance of the limb, comprising inertia, damping, and stiffness, provides a compact description of these stabilising properties. While inertia is determined by limb kinematics, stiffness and damping can be modulated via muscle activation, allowing humans to adjust interaction dynamics proactively. Understanding and modelling this modulation is not only important for basic neuroscience but is also highly relevant for the design of variable‑impedance robotic controllers, adaptive rehabilitation devices, and human–robot collaboration strategies.

Research questions

• How can multi‑joint arm impedance be estimated accurately in natural, dynamic movements, capturing both intrinsic and reflexive components?
• How are stiffness and damping modulated in response to task demands, sensory uncertainty, or partner performance, and how can these insights be incorporated into predictive impedance models?
• How can neuromechanical knowledge be integrated into grey‑box, uncertainty‑aware learning models that combine physical principles with data‑driven adaptability?
• How can such models provide prediction confidence and identify where additional data or refined modelling is needed, enabling safer and more personalised interaction strategies?
• How can impedance‑based predictive models be embedded in adaptive robotic control frameworks for stable, responsive, and comfortable human–machine interaction in collaborative and assistive scenarios?

Approach

Our research combines precision impedance estimation, computational modelling, and model‑based control design.

• Experimental impedance estimation
  • Developed a disturbance‑based framework capable of delivering short, controlled perturbations during free multi‑joint arm movements.
  • Designed a single‑trial method to isolate and quantify fast, involuntary feedback components, enabling the separation of intrinsic and reflexive impedance contributions.
• Modelling impedance modulation
  • Represented human movement as a stochastic dynamical system, embedding impedance parameters that vary with task context.
  • Analysed transient impedance modulation strategies, such as the dynamic adjustment of muscle coactivation to balance stability, effort, and information acquisition in physical interaction.
  • Identified how humans regulate impedance in multi‑agent settings to integrate visual and haptic cues optimally.
• Impedance‑based predictive control
  • Developed Gaussian Process and grey‑box learning models of impedance behaviour that incorporate neuromechanical priors.
  • Used these models for feedforward‑plus‑feedback control in simulated and experimental human–robot interaction tasks.
  • Enabled adaptive robotic impedance tuning to match user needs, stabilise interaction, and improve comfort and safety.

Key results and achievements

• Designed and validated an experimental framework for accurate multi‑joint impedance estimation during unconstrained arm movements.
• Developed a single‑trial isolation method for intrinsic and reflexive feedback components.
• Demonstrated that muscle coactivation strategies are adapted to improve interaction stability and sensory integration with a partner.
• Created impedance‑based learning models that fuse neuromechanical insight with probabilistic prediction for adaptive control.
• Transferred estimated impedance models to human–robot interaction controllers, enabling stable, variable‑impedance behaviour in collaborative and assistive scenarios.