As the use of robotics becomes more widespread, the huge amount of vision data leads to a dramatic increase in data dimensionality. Although deep learning methods can effectively process these high-dimensional vision data. Due to the limitation of computational resources, some special scenarios still rely on traditional machine learning methods. However, these high-dimensional visual data lead to great challenges for traditional machine learning methods. Therefore, we propose a Lite Fireworks Algorithm with Fractal Dimension constraint for feature selection (LFWA+FD) and use it to solve the feature selection problem driven by robot vision. The "LFWA+FD" focuses on searching the ideal feature subset by simplifying the fireworks algorithm and constraining the dimensionality of selected features by fractal dimensionality, which in turn reduces the approximate features and reduces the noise in the original data to improve the accuracy of the model. The comparative experimental results of two publicly available datasets from UCI show that the proposed method can effectively select a subset of features useful for model inference and remove a large amount of noise noise present in the original data to improve the performance.