Public Parts

Public Parts is an AI-managed platform for autonomous communal living. To subvert neoliberal platforms that currently control the gig economy, with a resocialized stance on social housing, the platform utilizes automation and machine intelligence to produce buildings capable of spatial reconfiguration for housing cooperatives. We seek to discretize and decentralize the typical apartment to minimize fixed ownership and maximize access. Public Parts’ AI manages a gig-based construction of discrete parts, automated spatial reconfiguration, communal upkeep, and a domestic task pool which operates as socio-economic infrastructure. The result is an environment that adapts to the behaviour of its inhabiting community while providing gig-work opportunities.

The platform utilises an existing paradigm as its operational basis – the housing co-operative. Through this communally owned and self-managed model, the platform finds its tenants, that manage and operate the housing project for themselves. The architectural setup of the platform is based on a discrete system of both fixed and configurable parts. These components are designed to use off-the-shelf wood elements that can be easily assembled through simple instructions and use of industrial robot arms, planned to be conveniently fabricated by the future tenants. The fixed elements serve as the framework for the space that is inhabited – its periphery, core and structure. They are automatically organized computationally and assembled together with post-tensioning steel rods.

The configurable components divide and define the habitable spaces. They are regarded as configurable because, through the action and movement of robots, they can be rearranged to form multiple spatial configurations. The robots proposed for the project are deliberately inspired by Amazon’s Kiva robots, here used to address the logistical problem of moving the elements that define the spaces inhabited by the community. The components also store simple parcels that contain an array of recallable domestic items, ranging from personal belongings, communally owned furniture, and personally owned beds.

As the configurable components are arranged on the slabs of the building, tenants occupy the spaces that are formed in between them. The reconfiguration of these components, therefore, incurs in the reconfiguration of the spaces in each slab, including their sizes and connectivity to each other. As a fundamental component to manage the organization of these spaces, an artificial intelligence algorithm is proposed to conduct the rearrangements that are required to maintain the suitability of the interior of the building for its inhabitants.

Spaces are occupied by the tenants by two means. Either they walk into a space and use it, or they utilise the platform’s application to request one. This request is processed by the Matchmaker, a facet of the AI that suggests a space according to the type of the activity and the number of people that will participate. In both methods, the AI utilises Bluetooth low energy beacons to detect the tenant’s position and automatically assign the space to them. After the period of occupation, individuals provide feedback based on their experience.

The strategy to achieve the emergent reconfiguration of the spaces is based on a continuous learning relationship between the tenants and the AI. To accomplish this, a machine learning method is implemented to engage in a feedback cycle with the inhabiting community. As tenants occupy the spaces in the building, they are asked to evaluate them, providing data that informs the AI and directs its reconfiguration orders. This interaction is done through the platform’s smartphone application, designed to be the interface between the AI and the inhabitants.

To operate on the interior arrangement of Public Part’s buildings, the AI must be able to recognize what constitutes a space in between configurable components. It was decided to implement a machine-learning algorithm to assist the AI in perceiving spaces in a manner like our own. Specifically, a customized implementation of pix2pix – an image-to-image translation Conditional Generative Adversarial Network.

The strategy consisted of translating the organization of a set of configurable components in a slab into an image, then utilising the pix2pix model to generate an output image including the separation between the spaces. Ultimately, this image is employed to define each space. The first step in this process was creating the training dataset, comprised of image pairs: one input image and another representing its expected translated output.

To produce the reconfiguration of the spaces, another artificial intelligence strategy is added to the project. The configurable components are defined as machine learning agents, that is, virtual entities that can take actions and make the observations necessary to inform such operations. These agents are the virtual representation of the parts, existing and operating in the AI’s voxel grid image of the building. In this environment, they take discrete actions to move and rotate, while also observing their position and orientation, the spaces they define, their proximity to other building elements, and the reconfiguration request they are meant to fulfil. Furthermore, these agents rely on a single “behaviour brain” to make the correlation between the observations and the appropriate actions to be taken

This brain represents the actual machine learning model that drives reconfiguration. It is subjected to reinforcement learning, i.e. by rewarding the model when proper actions are taken or objectives are satisfied, and applying penalties when incurring in incorrect results, or non-desired actions are taken. The model is trained by being exposed to several scenarios, in which its actions are supposed to modify the configuration and satisfy a random reconfiguration request. After training, the model will be able to respond to the different necessary reconfigurations the community might demand. As a result of the proposed learning and reconfiguring logic, any and every spatial configuration is possible, given that the geometrical constraints of the project are maintained.

Throughout the research, it was attempted to maintain a critical stance on the application of artificial intelligence in the branches of the architectural discipline. By investigating its technological definitions and potentials, it was sought to demystify the paradigm and envision how its combination with the built environment would influence architecture and its inhabitants.

Videos

Links

Bartlett b-pro show 2020
Bartlett b-pro show 2020 booklet

Project details: Design thesis developed for the MArch Architectural Design, b-pro programme, at the Bartlett School of Architecture [2019-2020].
Teammates: Keshav Ramaswami, Kiko Zhang
Tutors: Mollie Claypool, Gilles Retsin, Manuel Jimenez Garcia, Kevin Saey, Sonia Magdziarz, Alessandro Bava