Optimal Strategy for Sit-to-Stand Movement Using Reinforcement Learning

Background: Sit-to-stand motion is a frequent and challenging task in daily life activities especially for elderly and disabled people. Central nervous system uses several strategies for sit-to-stand movement. Many studies have been conducted to understand the underlying basis of the optimal approach. Reinforcement learning (RL) is a suitable method for modeling the control strategies that occur in neuro-musculoskeletal system. Methods: In this paper a dynamic model of human sit-to-stand was derived, and kinematic data of a healthy subject has been extracted in this task. An optimal control problem was formulated considering minimum energy and Q-Learning method has been utilized to find the optimal joint moments during sit to stand movement. Results: The simulation results have been compared to the experimental data. The lower extremity joint angles have been simulated and tracked the actual human angles extracted from the experiments. Also the joints moments showed a satisfactory precision by the proposed approach. Conclusion: An RL-based algorithm was used to model the human sit-to-stand, in which the model explores the state space with a Markov based approach and finds the best actions (joint moments) at each state (posture). In this approach the model successfully performs the task while consuming minimum energy. This was achieved by updating the algorithm in every trial using a Q-learning method