Assistive standing seat based on reinforcement learning

Sit-to-Stand (STS) is a key factor affecting the independent living of the elderly. Many older adults who can walk independently have to rely on the help of others because of the lack of independent STS ability. The Robot Assisted Standing Seat (RASS) can assist patients with tasks such as autonomous standing and post-operative rehabilitation. Under human biomechanics, RASS finds a compromise between task endpoint accuracy, body balance, energy consumption, and smoothness of motion and control. However, this method based on mathematical modeling requires more work to design a satisfactory control system. In addition, the specificity of the physical function of the elderly leads to great differences in their satisfaction with the RASS work process. This has aroused people's research on customized RASS to meet the needs of different users. To introduce user satisfaction into RASS, this paper proposes a deep reinforcement learning-based RASS. The proposed method takes the user's satisfaction as the reward function of the RL agent and trains a more reasonable control strategy according to the user's habits through online training. In this paper, we preliminarily verified the effectiveness of this idea in a simulation environment.