Learning Hip Exoskeleton Control Policy via Predictive Neuromusculoskeletal Simulation

Ilseung Park, Changseob Song, Inseung Kang

March 2026

TL;DR Train a hip-exoskeleton assistance policy entirely in a muscle-driven human walking simulation, with no motion capture, then distill it into a thigh-IMU controller on real hardware. In simulation the assistance cuts muscle activation by up to 3.4% and positive joint power by up to 7.0%, and on hardware the torque profiles match simulation with r = 0.82.

What they do

Exoskeleton controllers that work across speeds and slopes are usually trained on motion-capture and force-plate data. This paper instead trains a hip-exoskeleton policy entirely in a predictive neuromusculoskeletal walking simulation, with no motion-capture demonstrations. It then distills the policy into an IMU-only controller for a custom hip exoskeleton.

How it works

Simulation. A 21-DoF, 90-muscle human model in Hyfydy/SCONE, with bilateral hip torque actuators (up to 12 Nm) and 4.3 kg of added device mass.
Teacher. SAC with full simulator state. It acts through 10 muscle-synergy coefficients per leg (from NMF on human walking data), 10 trunk activations and 2 hip torques, at 0.7-1.5 m/s and slopes of -5° to +5°.
Curriculum. The model first learns to walk unassisted for 5×10^7 steps. Training then continues in two matched branches, with the exoskeleton and with it clamped to zero.
Student. A temporal convolutional network maps 0.95 s of thigh mediolateral gyroscope signal to hip torque. It is trained by behavioral cloning and runs at 100 Hz on a Jetson Orin Nano.

Results

Simulated gait matches open-source human data with mean RMSE 7.32° for joint angles and 0.264 Nm/kg for moments.
In simulation, assistance lowers muscle activation by 1.1% (level) and 3.4% (ramp ascent), and positive joint power by 6.3% and 7.0%. Ramp descent benefits less, and the reduction grows with speed.
With five treadmill participants, hardware torque waveforms match simulation with r = 0.82 ± 0.19 and RMSE 0.03 Nm/kg. Agreement ranges from r = 0.66 on ramp descent to 0.98 on ramp ascent.
Torque peaks lag the biological hip moment by 103 ms (extension) and 166 ms (flexion). These delays emerge from training instead of being hand-set.

Limitations

The human study was not an efficacy test: only torque and ground reaction forces were logged, so it does not show reduced effort for real users. Only cyclic walking is covered. A single gyroscope axis may not be enough for broader tasks and may be sensitive to sensor placement. Domain randomization was not tuned for sim-to-real.

Figures

From Figure 1: The sim-to-real workflow. A privileged teacher policy (SAC) is trained in a neuromusculoskeletal simul... — From Figure 1: The sim-to-real workflow. A privileged teacher policy (SAC) is trained in a neuromusculoskeletal simulation at 0.7-1.5 m/s and slopes of -5 to +5 degrees, first without and then with hip-exoskeleton torques, acting through a muscle-synergy prior. It is then distilled into a student network that maps thigh IMU gyroscope history to hip torque on the real exoskeleton.

From Figure 2: The simulated human-exoskeleton model: the H2190 musculoskeletal model (21 degrees of freedom, 90 Hill... — From Figure 2: The simulated human-exoskeleton model: the H2190 musculoskeletal model (21 degrees of freedom, 90 Hill-type muscles) with bilateral hip actuators applying flexion-extension torques.

From Figure 9: Change in mean muscle activation (top) and mean positive joint power (bottom) with assistance, relativ... — From Figure 9: Change in mean muscle activation (top) and mean positive joint power (bottom) with assistance, relative to no exoskeleton, as a function of walking speed for ramp descent, level ground and ramp ascent.

From Figure 10: Right hip assistance torque over the gait cycle in simulation (blue) and from the distilled student o... — From Figure 10: Right hip assistance torque over the gait cycle in simulation (blue) and from the distilled student on the real device (red), for ramp descent, level ground and ramp ascent at 0.7 to 1.5 m/s. Darker curves are faster speeds.

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, RMSE: 0.03 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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong