Understanding and reasoning in complex scenes are regarded as key indicators for evaluating the intelligence level of robots. However, existing methods mainly focus on question answering tasks in static environments and are insufficient for addressing multi-step reasoning requirements in dynamic scenes. A robot scene understanding and reasoning method based on distributed knowledge encoding was proposed. A scene knowledge graph with a dual-layer structure of concepts and attributes was constructed to enhance scene understanding capability. Entity concepts were represented as attribute sets through distributed embedding techniques, which enables compact and computable semantic representation. In addition, a classifier was designed to infer the sequence of robot actions required to achieve scene transformation based on the initial scene and the target scene. The proposed distributed knowledge encoding method exhibits good interpretability. Multi-step scene transformation processes were decomposed into executable single-step action sequences, such that the reasoning complexity increases linearly with the number of transformation steps. Experiments were conducted on the public Trance dataset. The experimental results demonstrate that, compared with GPT-4o, the proposed method achieves superior performance in scene transformation sensitivity and multi-step reasoning capability. Furthermore, feasibility and robustness in real-world scenarios are validated on a real robotic platform.