Learning Hip Exoskeleton Control Policy via Predictive Neuromusculoskeletal Simulation

Ilseung Park, Changseob Song, Inseung Kang
MetaMobility Lab, Carnegie Mellon University
arXiv

Real-time deployment of the learned hip exoskeleton controller across multiple terrains and walking speeds. The video shows hardware experiments together with the corresponding simulation, while the exoskeleton assistance torque is visualized in real time.

Abstract

Developing exoskeleton controllers that generalize across diverse locomotor conditions typically requires extensive motion-capture data and biomechanical labeling, limiting scalability beyond instrumented laboratory settings. Here, we present a physics-based neuromusculoskeletal learning framework that trains a hip-exoskeleton control policy entirely in simulation, without motion-capture demonstrations, and deploys it on hardware via policy distillation.

A reinforcement learning teacher policy is trained using a muscle-synergy action prior over a wide range of walking speeds and slopes through a two-stage curriculum, enabling direct comparison between assisted and no-exoskeleton conditions. In simulation, exoskeleton assistance reduces mean muscle activation by up to 3.4% and mean positive joint power by up to 7.0% on level ground and ramp ascent, with benefits increasing systematically with walking speed.

On hardware, the assistance profiles learned in simulation are preserved across matched speed-slope conditions (r: 0.82 ± 0.19, RMSE: 0.03 ± 0.01 Nm/kg), providing quantitative evidence of sim-to-real transfer without additional hardware tuning. These results demonstrate that physics-based neuromusculoskeletal simulation can serve as a practical and scalable foundation for exoskeleton controller development, substantially reducing experimental burden during the design phase.

From Neuromusculoskeletal Simulation to
Wearable-Sensor Control

Overall framework

We first train a privileged teacher policy in predictive neuromusculoskeletal simulation across randomized walking speeds and slopes, then distill it into an IMU-only student controller for real-time onboard deployment. This creates a direct pipeline from physics-based simulation to wearable-sensor exoskeleton control.

Learned Assistance Recovers Established Torque Structure

Learned assistance profile compared with established controllers

The learned assistance profile is not arbitrary: it recovers a waveform structure and peak timing that are broadly consistent with established exoskeleton control paradigms [1, 2]. This suggests that the simulation-trained controller discovers a biomechanically meaningful assistance strategy rather than merely exploiting the simulator.

References:
[1] Franks et al., Wearable Technologies, 2021.
[2] Molinaro et al., Nature, 2024.

Simulation-Learned Assistance Transfers to Hardware

Simulation versus real hardware

The assistance strategy learned in simulation is largely preserved on the physical device. Across matched speed–slope conditions, hardware torque profiles closely follow the simulated profiles, demonstrating sim-to-real transfer. The controller performs reliably on level ground and ramp ascent, but fails to provide effective assistance during ramp descent, suggesting that a sim-to-real gap still remains for certain locomotion conditions.

BibTeX

@article{park2026learning,
  title={Learning Hip Exoskeleton Control Policy via Predictive Neuromusculoskeletal Simulation},
  author={Park, Ilseung and Song, Changseob and Kang, Inseung},
  journal={arXiv preprint arXiv:2603.04166},
  year={2026}
}