This paper combines game decision-making and learning decision-making models to learn all possible types of strategies in robot behavior, describes human joints as a tree diagram structure through pose estimation, uses dynamic programming algorithms to derive joint information, extracts and models robot behavioral pose features, and identifies the action behaviors of construction robots. Through path tracking and other controls, robot behavior can be controlled to achieve the effects of construction robot behavior prediction and cooperative control. Set up simulation experiments to collect and preprocess the behavioral data of the construction robot, identify its behavior, predict its intent, and assess the safety risk of the construction robot’s action route. The robots constructed in this paper are put into the project, and the safety management input calculates the safety management efficiency. After adding the loss function to the model, the precision, recall, and F1 value mean of the construction robot are improved by 5.895, 5.461, and 5.765, respectively, and the derived safety management efficiency of the construction robot construction is 70, and the input of the construction robot brings a higher level of safety management to the construction project.