Robust neural network controller for a lower limb exoskeleton with minimal sensor configuration: a deep reinforcement learning approach with policy distillation.

Everything was tested in a computer simulation. No real person wore the device, so the authors' claim that it transfers to hardware is untested here. The modelled user had no voluntary muscle activity, standing in for quadriplegia. Spasticity, contractures and one-sided weakness, which are common in the patients you treat, were not modelled. Real-world sensor problems such as noise, drift and failure were not examined. The exoskeleton modelled weighs 20.4 kg, which is heavy for a device intended for home use. The authors designed and trained the controller and report their own results, with no independent comparison against an existing clinical exoskeleton. The paper never states how many trials were run for the push tests, so the 100% success rate at 40 N has no denominator.

The work was funded by Embry-Riddle Aeronautical University, the National Institute on Disability, Independent Living, and Rehabilitation Research, the New Jersey Health Foundation and the National Science Foundation. No commercial exoskeleton company is listed as a funder, and the authors report no conflicts of interest.

Do not read the 100% success rate under pushes up to 40 N as evidence this exoskeleton is safe for patients. It is a software model of a body with no muscle activation, never tested on a person, and the authors state real-world validation is still needed.