This paper presents a method to understand and reason the spatial configuration of individual parts in an assembled model from a 2D ISO sketch for autonomous small-part assembly. When given disassembled individual parts for a model in a 3D environment and an isometric 2D sketch of the assembled model, a human can tell visually the spatial configuration (position and orientation vectors) of the parts according to the assembly sketch, but it is challenging for a robot to decipher it. To understand the isometric drawings, a new reinforcement learning architecture is designed, which takes in an isometric drawing sketch as the reference and learns to reconfigure the spatial configuration of the parts to minimize the difference from the reference. A learning environment framework was created based on FreeCAD for reinforcement learning (RL) agent to interact with 3D models. The difference between the reference and current isometric sketches is calculated in the reward function, which guides the RL agent to move the 3D models to the target configurations. The proximal policy optimization (PPO) function was used for the RL agent because of its reliability and high training speed. To further augment generalization and robustness of the PPO model, an LSTM model was introduced into the RL architecture with a continuous action space. The proposed architecture demonstrated convincing performance in the experiment in FreeCAD. The performance indices include convergence on standard industrial assembly images within an average of about 60 steps, generalization on different initial configurations, and robustness to subtle nuances in a given reference sketch..