Robust Data-Driven Trajectory Generation and Control for Human-Like Assistance of an Upper-Limb Rehabilitation Robot

Abstract

Multi-degree-of-freedom trajectory generation is an important problem in human-robot interaction. While recent learning-based approaches have been developed to generate diverse trajectories, challenges still remain as applications involve human factors, object-oriented constraints, and heuristic movement strategies of human users. To address this issue, this study proposes a guidance strategy that incorporates the temporal structure observed in natural upper-limb motion. Using motion data collected while wearing the robot, an LSTMbased trajectory generator is developed and integrated with an Assist-As-Needed (AAN) strategy to form a guidance framework that adapts intervention according to task progress. The proposed method was implemented on a four-degree-of-freedom upper-limb rehabilitation robot and validated across various movement tasks. Experimental results show stable trajectory reproduction with a mean absolute error (MAE) of 0.62 cm and adaptive intervention under delay and disturbance conditions. These results highlight the potential of guidance strategies that consider both temporal motion characteristics and user state.