This document discusses using deep reinforcement learning for real-time assembly planning in robot-based prefabricated construction. It proposes a framework using DRL to help with planning issues like flexibility, uncertainties, and obstacle avoidance. A DRL model is presented using states, actions, and rewards to learn optimal assembly strategies through simulation. Experimental case studies test the approach in simulations of increasing complexity. Results show DQN and DDQN outperform other algorithms, with future work focusing on more realistic simulations and improved scalability.