The study firstly introduces the reinforcement learning theory, and proposes a decision-making method based on reinforcement learning to build a robot for autistic children, centered on autonomous human-robot interaction, with the purpose of serving the task of concentration training for autistic children. Among them, the goal task in the current environment is formulated based on imitation learning in the high level, and the robot’s action selection is realized based on interactive Qlearning in the low level. The decision making based on reinforcement learning to build a robot is applied to train the robot to interact with the training, and the simulation results verify the effectiveness and generalization of the designed algorithm in solving the concentration training path. Using the KANO model to analyze the needs of autistic children, based on which we design a multimodal human-computer interaction system for autistic children’s concentration training, and carry out a personal concentration intervention containing academic tasks for an 8-year-old autistic child, to verify the effectiveness of the multimodal human-computer interaction system in intervening in the concentration behaviors of autistic children, and the results of the study show that: the children’s concentration behaviors of the academic tasks in the intervention period are significantly improved compared with the baseline period compared with the baseline, and the mean value increased to 88.42%.