We introduce an approach to building a custom model fromthe on-the-shelf self-supervised models via their associating instead oftraining and fine-tuning. We demonstrate it with an example of a humanoidrobot looking at the mirror and learning to detect the 3D poseof its own body from the image it perceives. In order to build our model,we first obtain features from the visual input and the postures of therobot’s body via existing state-of-the-art models. Then we map theircorresponding latent spaces by a sample-efficient robot’s self-explorationat the mirror. In this way, the robot builds the solicited 3D pose detectorat one instant, instead of acquiring it gradually. The mapping, whichemploys associating the pairs of feature vectors, is then implemented inthe same way as the keys–value mechanism of the famous transformermodels. Finally, deploying our model for imitation to a simulated robotallows us to study, tune up and systematically evaluate its hyperparameters,without the involvement of the human counterpart, advancing ourprevious research.