SSM-Based Data-Driven Control for Soft Robotic Muscles

Data-driven slow-manifold modeling and inverse-SSM control for antagonistic HASEL-clutch joints (arXiv:2601.03247).

Outcome

I built a data-driven slow-manifold (SSM-based) inverse-model controller for antagonistic HASEL-clutch joints and achieved substantially lower tracking error than feedback-only and feedforward-only baselines.

Problem

Electrohydraulic artificial muscles have strong nonlinearities, hysteresis, and memory effects that make high-fidelity physics modeling expensive and hard to deploy in real time. This project learned a compact reduced-order model directly from forced-response data and used it for fast closed-loop control on hardware.

System

  • Hardware: antagonistic HASEL muscles with electrostatic clutches and joint-angle sensing
  • Modeling: spectral-submanifold reduction in the adiabatic regime, learning a slow-manifold map from forced-response trajectories
  • Control: inverse SSM feedforward plus PI feedback with anti-windup, saturation, and slew-rate limits
  • Validation: randomized unseen trajectory tracking under identical safety constraints across controller baselines

Contribution

  • Co-first author (equal contribution)
  • Data-driven model-reduction and control-pipeline design
  • Controller benchmarking design (feedforward-only, feedback-only, combined)
  • Experimental validation on antagonistic muscle hardware

Technical Stack

  • Nonlinear dynamics and spectral-submanifold reduction
  • Data-driven reduced-order modeling from forced-response trajectories
  • Inverse-model feedforward + PI feedback control
  • Real-time antagonistic soft-robot control experiments

Key Results

  • Combined inverse-SSM + PI controller: RMS error 2.38°, max error 10.41°
  • Feedforward-only baseline: RMS error 3.63°, max error 16.67°
  • Feedback-only baseline: RMS error 7.63°, max error 26.00°
  • Relative to feedback-only, combined control reduced RMS error by ~69% and max error by ~60%
  • Relative to feedforward-only, combined control reduced RMS error by ~34% and max error by ~38%

Media

Skills

nonlinear modeling data-driven control soft robotics experimental benchmarking