Reproduce “Learning Hip Exoskeleton Control Policy via Predictive Neuromusculoskeletal Simulation” on OpenResearch

About this paper

Ilseung Park, Changseob Song, Inseung Kang

Researchers at Carnegie Mellon University developed a sim-to-real learning pipeline using predictive neuromusculoskeletal simulation to train a hip exoskeleton control policy. The policy, distilled for physical hardware, reduced simulated user effort metrics and achieved average sim-to-real torque profile agreement of r=0.82 extpm 0.19.

Key insights

• Physics-based neuromusculoskeletal simulation can effectively front-load exoskeleton controller design, significantly reducing the experimental burden and accelerating development.
• The learned assistance policy from simulation successfully transfers to physical hardware, with quantitative preservation of assistance torque waveforms, demonstrating the viability of the sim-to-real pipeline.
• Exoskeleton assistance learned in simulation can reduce user muscle activation and positive joint power, with benefits increasing systematically with walking speed, serving as a predictive screening tool.

Summary by alphaXiv.