PRINCIPAL INVESTIGATORS: Matias del Campo, Sandra Manninger
LOCATION: Ford Robotics Building, North Campus, University of Michigan
Architecture has rarely found points of intersection with the research conducted on artificial intelligence on a global scale1. Even today, the discussion of AI and architecture has barely started. Considering the enormous potentialities of this area of research regarding its application in architecture, it is more than strange that this has not been discussed in wider circles within the discipline. In recent years we have seen rapid development in the progressive methods emerging from AI research, resulting in applications that surround us continuously. Almost undetected, AI applications have seeped successfully into our daily life: voice recognition, ride-sharing apps, banking apps, face recognition, AI airline pilots, intelligent home devices, and more are already naturally ingrained into our environment. More are in the pipeline, reaching from AI-driven cars to farming with intelligent machines2. The possibilities of these methods will transform all areas of our daily life and will mutate the planet. Consider the example of farming with intelligent machines. No need anymore for giant monocultures of crops. Suppose machines recognize the different crop species in a quasi “natural” environment consisting of multiple plant species. In that case, they can become mechanical gatherers, roaming the landscapes, harvesting to feed the planet. The result will be the end of the monocultural farm as we know it, and the rise of synthetic natures to optimize the yield. There are several advantages to this method: pathogens cannot spread anymore like wildfire through a monoculture, the need for pesticides is reduced enormously – there are many more. To work, machines must learn to perceive and understand their environment and autonomously respond to external stimuli3. Enter the Robot Garden.
How to test a Robot
SPAN (Matias del Campo & Sandra Manninger) have been in touch with AI experts since the late 1990s when they first came in contact with the faculty of the OFAI, the Austrian Institute for Artificial Intelligence. One of the oldest of its kind, it was founded in 1969 as the “Gesellschaft fur Kybernetik.” These early meetings provided a basis for understanding this area of research; however, the Robotics Institute of the University of Michigan’s input turned out to be a game-changer. After almost a year of conversations and experiments, the Robotics Institute offered SPAN the chance to design the Robot Garden based on 2D to 3D Style transfer techniques geared towards architecture design. What is the Robot Garden, you ask? First and foremost, it is a testing facility for robots. Michigan Robotics has specialized in exploring the possibilities of bipedal robots: robots on two legs. Combining its expertise in machine vision and machine learning, the team around director Jessy Grizzle has made essential steps forward in developing bipedal robots. These robots are designed to operate in areas usually intended for humans, such as factories, and in uneven terrain – think of the farming example mentioned above. In order o test these abilities, a testing ground was conceived, right next to Robotics’ new facility, the University of Michigan Ford Robotics Building. The outline given to SPAN called for a Robot Garden that contains a set of different terrains, from sand to grass to gravel, to the rockface. Inclinations and steps were part of the catalog of desired features to interrogate the “last 50 feet”4 problem.
Posthuman Design is here.
One of the first meetings we had about the Robot Garden was entirely dedicated to the conversation of agency in the age of AI. In recent years SPAN has discussed in length aspects of agency in a posthuman environment. Just to clarify, when we talk about posthuman, we do not mean “the age after humans,” but instead consider it an epoch that abandons the idea of human supremacy in the arts and design, and rather understands how other agents (or actors) can start to mingle with our ideas about creativity and sensibility. The more we understood about AI, the more we saw how collaborative design models could emerge from this area. The design technique used to design the Robot Garden is entirely embedded in this novel environment of design. Together with Alexa Carlson, a Ph.D. candidate of Michigan Robotics, we developed a design method based on deep neural networks. Deep neural networks have prevailed within many fields over the last few decades, including machine vision and natural language processing, due to their incredible accuracy at extracting salient features from input data and using these representations of our environment to perform tasks. This accuracy is thanks in part to the rapid development of powerful graphics computing technology. Access to big data has evolved, such that datasets can now start to capture the enormous amount of visual variance that exists in the world.
Generative adversarial networks (GANs) came into existence in 2014 as a machine learning methodology devised by Ian Goodfellow5. Leon Gaty’s paper A Neural Algorithm of Artistic Style6 was published in 2015. SPAN’s experiments with the use of Neural Networks in architecture started around 20187 with the design of the Austrian Pavilion for the Dubai Expo, and the first building project utilizing neural style transfer (NST) as a design method was the Robot Garden in 2019
Big Data, AI and Architecture Design
A discussion of architecture and AI does require us to touch upon the discussion about big data. Big data allows neural networks to learn what information is inherent in the data and how to decipher it: it is not about “Don’t Sort: Search“8, but rather about how to crunch through this big data to extract the relevant information that allows informing a project. It is literally about processing data to reveal information. Or to put it this way: “Data is the new oil.”9 Why? Because, similarly to crude oil—which is almost useless in its unrefined state, but needs to be refined into gas, plastics, or chemicals in order to create a valuable commodity—raw data is pretty much inert, as it is illegible to the human mind: it needs to be broken down and analyzed in order to reveal valuable information. Yes, data and information are two distinctly different things. This is also what makes the use of neural networks so incredibly powerful. Possible facets in the application of neural networks in architecture range from site analysis, plan analysis, and improved methods of Building Information Modeling, to aspects of the ecologic, economic, and social impact of a project. The opportunities to reveal the underlying nature of a project are gigantic. Using this notion, the project Robot Garden made use of a massive amount of satellite images in order to create datasets that informed the distribution of different terrains on the given site. After various attempts—some successful, others less so—
we found the right balance between the weights in the algorithm to come up with something that was useable as a testing ground for the robots.