GPT-5.1 Shows Emergent Spatial Reasoning as Physical Robot Controller Without Embodiment TrainingGPT-5.1无具身训练担任物理机器人控制器，展现类世界模型的空间推理能力
S 1.7T11 sources1 个来源R7-research
In an arXiv paper (Spinelli & Martins, submitted 27 Jul 2026, arXiv:2607.23899), researchers used GPT-5.1 as the high-level controller of a physical mobile robot with no prior embodiment, no simulator training, and no sensorimotor experience, feeding it only low-resolution first-person images and a discrete action set to drive navigation and object-directed tasks such as locating and contacting a target toy.
Across multiple trials, GPT-5.1 displayed emergent behaviors consistent with an internal world model, including retaining short-term memory of an object's location after it left the camera's field of view.
World models underpin path planning and decision-making in autonomous driving, so the finding hints that general-purpose LLMs may develop transferable spatial and physical reasoning relevant to future AV control architectures.