Vision-language models (VLMs) have shown potential for robot navigation but encounter fundamental limitations: they lack persistent scene memory, offer limited spatial reasoning, and do not scale effectively with video duration for real-time application. We present VL-KnG, a Visual Scene Understanding system that tackles these challenges using spatiotemporal knowledge graph construction and computationally efficient query processing for navigation goal identification. Our approach processes video sequences in chunks utilizing modern VLMs, creates persistent knowledge graphs that maintain object identity over time, and enables explainable spatial reasoning through queryable graph structures. We also introduce WalkieKnowledge, a new benchmark with about 200 manually annotated questions across 8 diverse trajectories spanning approximately 100 minutes of video data, enabling fair comparison between structured approaches and general-purpose VLMs. Real-world deployment on a differential drive robot demonstrates practical applicability, with our method achieving 77.27% success rate and 76.92% answer accuracy, matching Gemini 2.5 Pro performance while providing explainable reasoning supported by the knowledge graph, computational efficiency for real-time deployment across different tasks, such as localization, navigation and planning. Code and dataset will be released after acceptance.

We introduce an approach to building a custom model from ready-made self-supervised models via their associating instead of training and fine-tuning. We demonstrate it with an example of a humanoid robot looking at the mirror and learning to detect the 3D pose of its own body from the image it perceives. To build our model, we first obtain features from the visual input and the postures of the robot's body via models prepared before the robot's operation. Then, we map their corresponding latent spaces by a sample-efficient robot's self-exploration at the mirror. In this way, the robot builds the solicited 3D pose detector, which quality is immediately perfect on the acquired samples instead of obtaining the quality gradually. The mapping, which employs associating the pairs of feature vectors, is then implemented in the same way as the key-value mechanism of the famous transformer models. Finally, deploying our model for imitation to a simulated robot allows us to study, tune up, and systematically evaluate its hyperparameters without the involvement of the human counterpart, advancing our previous research.