Abstract: Bionic hands are promising devices for assisting individuals with hand disabilities in rehabilitation robotics. Controlled primarily by bioelectrical signals such as myoelectricity and EEG, these hands can compensate for lost hand functions. However, developing model-based controllers for bionic hands is challenging and time-consuming due to varying control parameters and unknown application environments. To address these challenges, we propose a model-free approach using reinforcement learning (RL) for designing bionic hand controllers. Our method involves mimicking real human hand motion with the bionic hand and employing a human hand motion decomposition technique to learn complex motions from simpler ones. This approach signifcantly reduces the training time required. By utilizing real human hand motion data, we design a multidimensional sampling proximal policy optimization (PPO) algorithm that enables efcient motion control of the bionic hand. To validate the efectiveness of our approach, we compare it against advanced baseline methods. The results demonstrate the quick learning capabilities and high control success rate of our method.

Key words: Bionic hand , · Reinforcement learning , · Motion decomposition , · Multidimensional sampling PPO algorithm