Reinforcement Learning Control Outperforms Iterative Learning in Exoskeleton-Assisted Gait Training
Andy Li, Haoran Li, Aytac Teker, Mariana Hernandez-Rocha, Biruk Gebre, Karen Nolan J., Kishore Pochiraju, Damiano Zanotto
AI summary
Problem
Direct head-to-head comparisons of traditional adaptive versus modern learning-based assist-as-needed controllers for powered exoskeletons are lacking, obscuring whether performance gains stem from the control algorithm or hardware differences.
Approach
We performed a within-subject crossover study comparing an RL-based and an ILC-based assist-as-needed controller on a custom ankle exoskeleton, using a perturbed-gait protocol on a self-paced treadmill with healthy participants to measure stride velocity improvements and retention.
Key results
- Comparable immediate stride velocity gains during training for both controllers
- Greater adherence to target walking speed with RL-AAN during training
- Superior short-term retention of speed improvements post-training with RL-AAN
- Higher stride-to-stride control variability indicating dynamic adaptation with RL-AAN
Why it matters
These results validate reinforcement learning as a promising adaptive control strategy for personalized exoskeleton rehabilitation, guiding future clinical applications for gait impairment recovery.
Abstract
Learning-based controllers are increasingly adopted in lower-extremity powered exoskeletons, yet their advantages over traditional adaptive approaches remain underexplored. We compared two adaptive assist-as-needed (AAN) controllers for gait training with an ankle exoskeleton: a reinforcement learning-based controller (RL-AAN) and a conventional iterative learning controller (ILC-AAN). Both adjusted assistance stride-by-stride, delivering torque as a percentage of the wearer’s biological plantarflexion moment—estimated online with a subject-agnostic model—and progressively faded assistance as performance improved. Healthy participants walked on a self-paced treadmill under a perturbed-gait protocol. Performance was assessed as average percent stride-velocity (SV) improvement relative to unassisted perturbed walking (∆%ε+ SV ) and percent of strides above the SV threshold (N%+ SV ). During training, RL-AAN and ILC-AAN elicited comparable gains in ∆%ε+ SV , but RL-AAN yielded greater adherence, as indicated by larger N%+ SV . After training, RL-AAN demonstrated superior retention in ∆%ε+ SV and N%+ SV . These results support RL-AAN as a promising strategy for subject-tailored gait training, motivating future studies in clinical cohorts with neurological or musculoskeletal gait impairments.