Learning Ergonomic Control in Human-Robot Symbiotic Walking

This article presents an imitation learning strategy for extracting ergonomically safe control policies in physical human-robot interaction scenarios. The presented approach seeks to proactively reduce the risk of injuries and musculoskeletal disorders by anticipating the ergonomic effects of a robot's actions on a human partner, e.g., how the ankle angle of a prosthesis affects future knee torques of the user. To this end, we extend ensemble Bayesian interaction primitives to enable the prediction of latent biomechanical variables. This methodology yields a reactive control strategy, which we evaluate in an assisted walking task with a robotic lower limb prosthesis. Building upon the learned interaction primitives, we also present a model-predictive control (MPC) strategy that actively steers the human-robot interaction toward ergonomic and safe movement regimes. We compare the introduced control strategies and highlight the framework's ability to generate ergonomic, biomechanically safe assistive prosthetic control. A rich analysis of constrained MPC shows a 20× reduction in the effects of large perturbations on prosthetic control system. We empirically demonstrate a 16% reduction in vertical knee reaction forces in real-world jumping experiments utilizing our control methodology and examine other optimal control strategies in simulated walking experiments.