Tethered Architectures in Cyber-physical System Development: The Case of Tesla's Autopilot System

Cyber-physical systems form a complex design challenge that requires learning through product use. A central design objective is a contextually and physically embedded behavioral model that controls situated interaction with respect to environmental variation. Environmental variety and systemic complexity, however, bounds developers’ capacity to foresee all eventualities, calling for a process and architecture that facilitate physically grounded experimentation and adaptation. Current theories do not provide an adequate architectural blueprint. Building on a longitudinal analysis of Tesla's Autopilot system, we propose the concept of a “tethered architecture.” A tethered architecture accommodates product hardware's slower, incremental iteration to be linked with faster software development and continual data- rich feedback loops. It foregrounds the circular flow of data and code between consumer-facing product instantiations and backend software operations. It also discloses the multifaceted informational and computational connectedness which permeates across product instantiations and through the development process. A tethered architecture combines programming, machine learning, data production, testing, and simulation – highlighting how functional interdependencies and interaction effects shape data collection, feature iteration, and learning. We argue that cyber-physical systems require a shift from an architectural language of separation to one that embraces high levels of systems interconnectedness between the software and its enabling hardware.