DeepCanvas: Sequential design strategy acquisition in context-aware building footprint synthesis using vision-based deep reinforcement learning

Building footprint synthesis in architectural design is a complex task that involves constraint satisfaction and objective optimization, traditionally relying on human expertise. This paper presents DeepCanvas, a novel framework that integrates procedural modeling with vision-based deep reinforcement learning (deep RL) to generate context-aware building footprints. Our approach represents designs through sequential parametric manipulations on primitive shapes, including geometric transformations and Boolean operations. The framework combines neural networks' pattern recognition capabilities with reward-driven exploration, enabling the RL agent to optimize architectural criteria through direct interaction with the design space rather than from existing samples. Although currently experimented with basic spatial quality metrics and simplified primitives for computational efficiency, results demonstrate that our agent successfully discovers context-adaptive strategies for generating optimized layouts from a vast design space while maintaining constructive outputs. The framework bridges machine-learnable and human-manipulable representations in architectural design, offering a systematic approach to learning-based design optimization that preserves procedural modeling’s interpretability and editability.