The paper introduces the Functional Manipulation Benchmark (FMB), a real-world benchmark designed to study robotic learning in functional manipulation tasks. FMB emphasizes a balance between complexity and accessibility, ensuring that tasks are narrow enough to be effectively tracked with manageable datasets while being diverse enough to pose significant generalization challenges. The benchmark includes 3D-printed objects and a dataset of 22,500 human demonstrations, covering grasping, repositioning, and assembly skills. FMB is modular, allowing researchers to focus on specific stages or aspects of the task and evaluate different methods. The paper also presents an imitation learning framework with pre-trained policies for individual stages and multi-stage tasks, facilitating the study of various parts of the pipeline. The benchmark aims to serve as a versatile toolkit for researchers to explore robotic learning in functional manipulation tasks, addressing both generalization and physical complexity.The paper introduces the Functional Manipulation Benchmark (FMB), a real-world benchmark designed to study robotic learning in functional manipulation tasks. FMB emphasizes a balance between complexity and accessibility, ensuring that tasks are narrow enough to be effectively tracked with manageable datasets while being diverse enough to pose significant generalization challenges. The benchmark includes 3D-printed objects and a dataset of 22,500 human demonstrations, covering grasping, repositioning, and assembly skills. FMB is modular, allowing researchers to focus on specific stages or aspects of the task and evaluate different methods. The paper also presents an imitation learning framework with pre-trained policies for individual stages and multi-stage tasks, facilitating the study of various parts of the pipeline. The benchmark aims to serve as a versatile toolkit for researchers to explore robotic learning in functional manipulation tasks, addressing both generalization and physical complexity.