Predictive safety framework using Kalman smoothing and LSTMs for robotic exoskeletons

Abstract

In this paper, we develop a predictive safety control framework for robotic exoskeletons that integrates Kalman smoothing (KS) for noise reduction in sensor data, long short-term memory (LSTM)-based force prediction, adaptive impedance control, and dynamic barrier functions. The proposed approach is applicable to safety-critical tasks in uncertain environments and addresses important limitations of prior methods including disjointed perception control pipelines, rigid safety constraints, and lack of guaranteed stability. The main contributions of the proposed approach are: a unique KS-LSTM coupling to reduce interaction force prediction errors by 44.5% while operating in real-time (22.3 ms/cycle); adaptive control that modulates stiffness/damping parameters using LSTM-predicted forces; and context-aware barrier functions with formal globally uniformly ultimately bounded (GUUB) stability guarantees via Lyapunov analysis. Through extensive simulations, the proposed approach demonstrates superior performance over baselines in prediction accuracy, trajectory tracking, and constraint compliance (95.04%), especially in obstacle avoidance and human–robot collaboration. This framework connects learning-based prediction and real-time certifiable control, which helps to ensure safer physical human–robot interaction in rehabilitation, industrial, and assistive settings.