The paper introduces Language-Guided Abstraction (LGA), a method that uses natural language and pre-trained language models (LMs) to automatically construct state abstractions for imitation learning. LGA aims to improve the efficiency and effectiveness of policy learning in high-dimensional observation spaces by identifying and isolating task-relevant features. The process involves three main steps: textualization, feature abstraction, and instantiation. Textualization transforms raw perceptual inputs into a text-based feature set, feature abstraction uses an LLM to select relevant features based on the task description, and instantiation converts the abstracted features back into a perceptual input. The learned policy is then trained using a small number of demonstrations and the abstracted states generated by LGA. Experiments on simulated robotic tasks show that LGA produces state abstractions similar to those designed by humans but with significantly less human effort. These abstractions enhance the generalization and robustness of the learned policies, even in the presence of spurious correlations and ambiguous linguistic specifications. The utility of LGA is demonstrated on real-world mobile manipulation tasks using a Spot robot.The paper introduces Language-Guided Abstraction (LGA), a method that uses natural language and pre-trained language models (LMs) to automatically construct state abstractions for imitation learning. LGA aims to improve the efficiency and effectiveness of policy learning in high-dimensional observation spaces by identifying and isolating task-relevant features. The process involves three main steps: textualization, feature abstraction, and instantiation. Textualization transforms raw perceptual inputs into a text-based feature set, feature abstraction uses an LLM to select relevant features based on the task description, and instantiation converts the abstracted features back into a perceptual input. The learned policy is then trained using a small number of demonstrations and the abstracted states generated by LGA. Experiments on simulated robotic tasks show that LGA produces state abstractions similar to those designed by humans but with significantly less human effort. These abstractions enhance the generalization and robustness of the learned policies, even in the presence of spurious correlations and ambiguous linguistic specifications. The utility of LGA is demonstrated on real-world mobile manipulation tasks using a Spot robot.