Spatial AI

In 2022, OpenAI’s ChatGPT and DALL-E highlighted techniques called “generative artificial intelligence,” which have since captivated the world with remarkable creative abilities. Simple text prompts can now appear to write essays, perform deep research, reason deductively, generate striking imagery and video, even author and debug code.

At the same time, AI pervades the built environment. Government agencies track and identify cars on public roadways; autonomous vehicles and dog-like robots can continuously map environments in real time; environmental sensors record changes in air quality and climate; tech companies listen to and monitor our homes; and satellite data feed increasingly sophisticated predictive models of urban growth and destruction.

Spatial designers (architects, urban planners, etc.) face particularly complex problems when proposing changes to the built environment, a task which at best requires anticipating the consequences of such decisions. But current software tools don’t intrinsically carry the semantics of “space;” while useful, they rather focus on manipulating 3D geometries or forms of physical data.

Beyond the 3D and physical, “spatial AI” refers to AI as applied to spatial reasoning, the logic of navigating, designing, and operating in space. Such a medium would enable designers to focus on the semantics of their spatial design propositions and environments, rather than focusing on those of algorithms or commercial software.

This research seminar will explore the potential of 3D, physical, and generative AI to facilitate insights, decisions, and predictions for problems involving higher-level spatial reasoning.

In this course, students will:
• explore the definitions, affordances, and inner workings of generative and discriminative AI, scrutinize canonical writings from relevant technological, architectural, and computational theories, particularly regarding notions of “space,” • experiment with the rapidly evolving landscape of AI methods, including large language models, computer vision, robotics simulations, etc. • craft semantic models and spatial ontologies, interpreted by LLMs, • develop a critical and technical understanding of the technologies, and • speculate on new spatial AI methods at human, architectural, and urban scales.

New AI methods are introduced weekly using the current open platforms (primarily HuggingFace but also OpenAI and Google Vertex AI). Class sessions each involve a short lecture, an intensive workshop, and student presentations, with readings and technical prep work in-between sessions. The course culminates in a final project.