With the progress of science and technology, multifunctional rehabilitation robots have come into people's lives, so improving the robot's tactile perception has become crucial. This work first describes mechatronics and explains relevant technologies and their current applications in multifunctional rehabilitation robots. Then, a method of combining the neural network and a multifunctional rehabilitation robot's tactile sensor is proposed, which optimizes the sensor through neural network technology, thus improving tactile perception. Finally, Long Short-Term Memory (LSTM) is employed to further optimize the model and improve its stability. Comparative experiments are conducted to verify the model's effectiveness. The experimental results show that 25 iterations are the boundary value of the optimization model, regardless of training or testing. The learning speed of the first 25 iterations is relatively fast, and the loss of iteration starts to decline after the 25 iterations. The decline speed is relatively slow, but the accuracy has become higher. With the increase of iteration times, although the loss is still decreasing, the highest accuracy rate during the test has reached 97% and it tends towards stability. It proves the effectiveness and rationality of the optimized model. The soft and hard recognition of objects is distinguished through the second experiment. It suggests that when recognizing soft and hard objects, the average recognition accuracy of the model after optimization can reach 0.965, while the average recognition accuracy of the traditional model is only 0.915, which further verifies the feasibility of the optimized model. Therefore, this work has certain reference significance for optimizing multifunctional rehabilitation robots.