Automotive sensing technologies and the automation of space: The case of Subaru EyeSight

Abstract

While a great deal of critical attention has been focused on the experimental use of automotive sensing technologies within electronic vehicle development, their integration into production cars continues apace. Despite a long history of incorporating vision and other automated sensing systems into production cars, we are yet to fully appreciate how these ‘smart objects’ function and interoperate, how ‘observable’ these systems are (Rieder and Hofmann, 2020), what roles they play in shaping our day-to-day experiences of the automation of automotive space. In this article, we examine the vision-sensing capacities of one proprietary in-vehicle computer vision system (Subaru EyeSight) and how its visual sensors operate and interoperate to form a unified ADAS (advanced driver assistance system) accident mitigation system. In particular, we seek to understand the forms of driver-to-vehicle handover that modern ADAS facilitate and require. Drawing on social scientist James Ash’s concept of ‘phase space’, and cyberneticist W. Ross Ashby’s earlier conception of ‘phase-space’, our contention in this article is that ADAS like EyeSight actively intervenes in and reconfigures automotive space. We detail how EyeSight’s automated vision sensors and decision-making system reshapes how automotive space is viewed, understood, and acted upon, with implications for those behind the wheel.

Introduction

We may think of the incorporation of sensors into cars as a relatively recent phenomenon. In truth, cars have long relied upon sensor technology for their successful operation. Whether it is analogue or digital, the humble fuel gauge, for example, is connected to a simple sensor that determines the level of fuel in the tank and communicates this to the car dashboard display.

What has changed over time, however, is the sheer proliferation and sophistication of automotive sensors, and their coupling to automated decision-making systems. Modern cars are now packed full of sensors (and electronic actuators) performing a variety of monitoring, measurement, and operational functions. There are sensors for virtually all aspects of the car’s operation: engine, transmission, braking, fuel system, temperature, pressure, speed, oxygen, positioning, as well as performance monitoring, problem detection, and efficiency improvement. These sensors send information to the car’s computers – its electronic control units (ECUs) – which act on this information to optimize or adjust vehicle performance. Writing in 2004, Nigel Thrift observed how the car had become a ‘software platform’, with computing software an ‘integral element of the mechanics of cars’ that are available for mass consumption:

Now software controls engine management, brakes, suspension, wipers and lights, cruising and other speeds, parking manoeuvres, speech recognition systems, communication and entertainment, sound systems, security, heating and cooling, in-car navigation and, last but not least, a large number of crash protection systems. (Thrift, 2004a: 50)

There is no software without hardware. Twenty years on, and MC Forelle (2022) describes the intensification of the ‘car-driver-software assemblage’ (Featherstone, 2004: 10) as the ‘chipification’ of cars – the embedding of microchips into vehicles to facilitate sensor-driven software and data collection functions.

A key driver of this intensification of in-car computing technology is the deployment of sensors, chips, and software to facilitate automotive automation. As the Society of Automotive Engineers’ (2014) explanation of International Standard J3016 sets out, automotive automation occurs along a spectrum from Level 0 (no driving automation) to Level 5 (full driving automation). Of specific interest in the context of this article are sensing and computing technologies deployed in the operation of advanced driver assistance systems (ADAS) – the crash protection systems noted above by Thrift. In Australia, as of 2023, around 60% of new cars contained ADAS (AAAA, 2023), with these systems being progressively incorporated – and in many jurisdictions, mandated – in new models. The sensor technologies employed (in different configurations) within ADAS include cameras, LiDAR (Light Detection and Ranging), sonar, radar, ultrasonic sensors, GPS (Global Positioning System), and sometimes inertial navigation systems, consisting of gyroscopes, accelerometers, and magnetometers or GNSS (Global Navigation Satellite Systems) receivers.

Modern vehicles with increasingly sophisticated ADAS safety features – electronic stability control, auto emergency braking, traction control, driver attention detection, antilock braking system (ABS), collision warnings – generally fall within automation Level 1 (driver assistance) and Level 2 (partial automation) (Leiman, 2021: 225). And yet, the interventions these systems make to vehicle operation and driver control are substantial:

Although not Level, 3, 4, or 5 vehicles, [. . .] ADAS features may effectively override human drivers’ capacity to direct vehicle operation. Some prevent drivers from operating the vehicle at all, some become active when drivers lose control, others alert drivers to imminent risks, and yet others shut down the vehicle or other devices when being operated unsafely. (Leiman, 2021: 256)

A great deal of critical and public attention has been focused on development at the higher end of the SAE automation spectrum – so-called autonomous vehicles. All the while, sensing technologies are being increasingly integrated into mass-market production cars, facilitating specific forms of ADAS-related automated decision-making, while garnering far less critical and public attention. Despite a long history of incorporating sensing systems into production cars, we are still to fully appreciate how these ‘smart objects’ (Ash, 2019) function and interoperate, and what roles they play in shaping and reconfiguring our day-to-day experiences of the automation of automotive space. While regulatory and policy attention focuses on the digital media platforms we interact with on our smartphones, the systems we rely on for everyday mobility are far less well understood.

This article examines ADAS sensing technologies and the ways we may approach them. In this article, we will argue that contemporary production cars have become sensor-computational ‘smart objects’, exemplified by Subaru’s dual-camera EyeSight ADAS system, which actively reconfigures the experience and governance of automotive mobility through low-level automation. We will demonstrate how ADAS systems redistribute agency across drivers and machines, generating new spatial relations and regulatory mechanisms in real time, and propose a reframing of ADAS as a mode of spatial automation rather than merely a safety feature. By mobilizing the concept of phase space from geographer James Ash (2019) and from cyberneticist W. Ross Ashby (1956), the article builds on earlier geography scholarship on automotive automation and the production of space to reveal how ADAS and automated decision-making interfaces significantly reshape driver engagement with vehicle operation and movement-space. The article argument is structured as follows. We begin by reviewing existing understandings of contemporary automobiles as smart objects, before introducing the framing concept of ‘phase space’. From there we provide a condensed history of modern ADAS technologies, before exploring the development of Subaru’s proprietary system, EyeSight. With this as context, we detail the operation of the EyeSight system, and how it surfaces two key forms of data: vehicle data and driving data. We conclude the article by reflecting on what it means for ADAS systems such as EyeSight to create automotive space in real time.

Understanding contemporary automobiles as ‘smart objects’

Within communication and cognate fields, scholars have explored the technologization of cars along several paths. Contemporary connected cars have been defined as a species of ‘automated media’ (Andrejevic, 2020: 29); ¹ as ‘mobile spatial media’ (Alvarez León, 2019), or automotive ‘geomedia’ (Thielmann, 2007); as a ‘communication platform’ (Featherstone, 2004; Wilken and Thomas, 2019); and as an ‘automotive platform’ (Hind et al., 2022; Hind and Gekker, 2024), ² with the last adding a novel dimension to Urry’s (2004) idea of ‘the “system” of automobility’. Work on automotive digital platforms notes how these platforms depend on ‘new forms of data collection via diverse arrays of sensors from cameras and LIDAR to GPS’ and how such data streams are ‘enrolled in new mechanisms of location, valuation, and marketization’ (Alvarez León, 2024: 131). Recent scholarship also focuses specifically on the ‘sensor work’ that is performed by contemporary autonomous vehicles (Hind, 2023, 2022; Mukhtar-Landgren and Paulsson, 2023), and how, often, the ‘primary component that interfaces with the various sensors’ in advanced automobile systems is visual perception (McCosker and Wilken, 2020: 102) or computer vision (Dobson, 2023).

Contemporary connected vehicles have been the subject of intense interest within geography scholarship, with particular attention paid to how they reconfigure urban space and our experience of it. Amin and Thrift (2002) describe a reconstitution of everyday life ‘as a cloud of informational devices begins to descend over the city, bringing with them another informational remediation’ (p. 102). Software (or code) is regarded as a central feature of these ongoing transformations. Such is the impact of everyday software-encoded objects on urban life, Dodge and Kitchin (2005) argue, that they reconstitute space, leading to the creation of what they term ‘code/space’. This involves an understanding of space as ‘ontogenetic’, as something that it is ‘transduced’, with space continuously brought into being through practices of people in combination with software (Kitchin and Dodge, 2011: 71–73). Software-encoded technologies are significant in geographical terms, they argue, as they ‘modulate the form, function and meaning of space – they produce space in new ways’ (Dodge and Kitchin, 2007: 271). They do this by generating ‘new spatialities’ and by creating ‘software-sorted or machine readable geographies’ (2007: 264).

In a related vein, Thrift (2004b: 592) draws attention to the new and important forms of calculation that accompany software-encoded sensor systems. He writes, ‘the sheer amount of calculation that is now becoming possible at all points of so many spaces is producing a new calculative sense’. Thrift refers to this, following the work of Michel Callon and John Law (2005), as ‘qualculation’ – the mutual consideration of the quantitative and qualitative aspects of decision-making. Thrift details the manifold things that this ‘new calculative sense’ relies on. This list includes the following: ‘a series of prostheses which routinely offer cognitive assistance and which do much of the work of navigation automatically’; ‘a highly provisional sense of spatial co-ordination which is based in [. . .] continual spatial and temporal revisions’; and ‘continual access to information [. . .] arising out of connectivity being embedded in all manner of objects’ (Thrift, 2004b: 593). Thrift’s account of ‘qualculation’ is prescient in that it can be read from the contemporary conjuncture as describing modern-day ADAS technologies almost to a tee.

Indeed, the rise of contemporary connected and automated vehicles, with their sensing, data processing, and information transmission capacities are viewed in this literature as ‘inseparable’ (Sheller, 2007: 186) from these understandings of ‘code-written spaces’ (Thrift and French, 2002) and from the generation of a ‘new regime of movement-space’ and revised forms of spatial calculation (Thrift, 2004b: 595). Modern automobiles are ‘stuffed full of software’ (Thrift and French, 2002: 313) and are seen to represent ‘one of the densest concentrations of digital computing embedded software that most people encounter in their everyday environment’ (Dodge and Kitchin, 2007: 267). These vehicle-embedded computing technologies permit new means of spatial calculation and spatial processing that generate new forms of ‘automated spatialities’ (Merriman, 2009: 590) with implications for the ‘wider movement-space of automobility’ (Sheller, 2007: 178).

In this article, we extend this prior work through an investigation of the vision-sensing capacities of one proprietary in-vehicle computer vision system (Subaru EyeSight) and how its visual sensors operate and interoperate to form a unified ADAS accident mitigation system. In particular, we seek to understand the forms of driver-to-vehicle handover (Goldenfein et al., 2020; Runciman, 2023) that modern ADAS facilitate and require. Beyond ADAS providing drivers with ‘varying degrees of computationally supported driving aides’ (Forelle, 2022: 5), our contention in this article is that EyeSight actively intervenes in and reconfigures automotive space (understood here as the space in and around a moving vehicle). As we will go on to detail in this article, EyeSight’s automated vision sensors and decision-making system reshape how automotive space is viewed, understood, and acted upon, with implications for those behind the wheel.

Our investigation of the question of how ADAS technology reconfigures automotive space is informed by a lineage of work on the new calculative sense and transduction of ‘movement-space’ through software-mediated practices. We draw further inspiration from geographer James Ash’s theorization of smart objects. In his book Phase Media, Ash (2019) describes smart objects as sensor-enabled devices that alter a user’s experience of their environment. Ash explores how smart objects possess both intentionality and protentiality. With respect to intentionality, Ash’s argument is that the combination of sensor components within smart objects gives them ‘differing levels of intentionality that is key to [. . .] what smart objects can do’ (Ash, 2019: 37). Smart objects possess forms of intentionality in that they are designed and manufactured with various sensors that ‘enable a specific form of “towardness” to other things’ (p. 38). These orientating intentionalities are generated ‘both to furnish smart objects with particular functions and to allow the companies who create these objects to control how they are used’ (p. 38). Protentiality, then, is the capacity to anticipate what comes next. Ash argues that ‘the protentiality particular to smart objects can be defined as their ability to await specific forces or signals in the environment and respond to these forces and signals through the way they perturb or are perturbed by other things’ (pp. 46–47). In the case of automobiles, we might think of how camera sensors ‘perturb’ the CPUs within the car, or how a radar sensor detects a vehicle approaching on the left side of the car and ‘perturbs’ the blind spot notification sensor.

Ash combines these ideas in his concept of ‘phase space’. Drawing from Simondon’s understanding of ‘phase’, where ‘the term “phase” refers to how objects emerge and change within a milieu’ (Ash, 2019: 59), Ash employs phase ‘specifically to emphasize the temporary yet successive and overlapping nature of the spaces smart objects generate’ (p. 59). ‘Phase spaces’, Ash suggests, ‘can be defined as the relations of proximity and distance produced by the qualities disclosed by smart objects as they intentionally and protentionally perturb and are perturbed by other smart objects and objects in general’ (p. 60).

According to Ash, how humans and smart objects experience space are two quite different things. This has implications for how we as humans engage and interact with smart objects, such as production cars and the automated sensor systems that are embedded within them. Ash (2019) writes,

When we recognize that smart objects perturb one another alongside or outside the purview of human experience, it quickly becomes clear that the variety of smart, technical and other non-technical objects present at any one moment can create multiple phase spaces. (p. 60)

Ash explores this idea of phase space – or multiple phase spaces – in relation to a range of technologies (the iPhone, Nest Cam, and weather and running apps), while also acknowledging its wider applicability, including cars. Writing on autonomous vehicles, like the Waymo car for instance, Ash notes that it ‘intentionally perturbs objects using a range of sensors and cameras in order to create an intelligibility of space’ (p. 74).

In this article, we draw on Ash’s formulation of the phase space in examining ADAS-equipped cars as smart objects that reshape the intelligibility of automotive space for driver and passenger. We note, however, that there is another, much earlier, conceptualization of phase spaces which appeared in a distinct but orthogonal domain of thinking about technology, in W. Ross Ashby’s An Introduction to Cybernetics (Ashby, 1956: 37–39). It is curious that Ash makes no mention of his near namesake’s work, even taking into account the disciplinary tensions – manifest from the 1950s – between Simondon’s philosophy of technology and Weiner’s cybernetics. Ashby’s phase-space is not Ash’s phase space: rather than describing how technologies shape people’s experience of space and time, the Ashby phase-space offers a model for representing all the relevant possible states of a system, highlighting the degree to which the multiple components in a system change state, and how the state of a system overall may adapt, and what may be required to regulate and stabilize it: ‘the whole range of trajectories in the system can be seen at a glance, frozen, as it were, into a single display’ (Ashby, 1956: 38).

The differences between these two conceptions of the phase space should not be underestimated. In this article, we are not concerned with revisiting, or attempting to reconcile, old arguments over the technical ontologies of cybernetics, on the one hand, and the putative reduction, or privileging, of human experience, on the other. It is significant, however, that Ash’s formulation is explicitly not restricted to the domain of human technological experience or human interaction with smart objects; it is also intended to encompass the changing states – ‘perturbations’ – of the various elements comprising smart objects. At the same time, Ashby’s phase-space is explicitly representational; it is designed to address precisely the questions of visibility, overall control, systemic regulation, and stability that are central to automated automotive sensing. There are therefore connections as well as divergences between these different conceptual frames, even if these are unintended and unrecognized. We suggest that these points of connection may prove useful in understanding automotive phase spaces.

Before we turn to these concerns and the case of Subaru EyeSight, we offer some historical context, to show how these driver assistance systems have their own trajectories, and how their combination and integration over recent decades has proved to be a critical development.

Histories of driver assistance systems

ADAS technologies, such as those found within contemporary production cars, have been recently named and organized under this label, but they did not arrive fully formed. The various elements of these systems – such as radar, cruise control, and rear-facing and forward-facing cameras, each of them addressed in turn below – have their own histories of development and technological refinement. For instance, radar first appeared in 1959 in the futuristic Cadillac Cyclone concept car, with sensors for crash avoidance housed in two front-mounted Dagmar nose cones (Bowman, n.d.) (see Figure 1). Interest in the technical possibilities of automotive radar for collision prevention continued throughout the 1960s and 1970s (Meinel and Dickmann, 2013; Merlo, 1964). Use of radar, sonar, and LiDAR are now common inclusions within many contemporary vehicle avoidance systems.

Figure 1.

Precursors to modern-day ADAS: the 1959 Cadillac Cyclone concept car, with nose cone housed radar sensors for crash avoidance.

Cruise control first appeared, in mechanical form, in a Wilson-Pilcher car from 1904, where the driver activated a lever on the steering column to set the cruising speed (Bober, 2016). However, cruise control as we now know it was developed by noted US automotive engineer and inventor Ralph Teetor in the 1940s, and became standard issue in Cadillac cars by the late 1950s (Bober, 2016). More recently, adaptive cruise control has added radar (which Mercedes Benz did with their early DISTRONIC system, from 1998) and cameras to regulate the speed of one car in relation to another ahead on the road (Meinel and Dickmann, 2013; Waldschmidt et al., 2021). Forward-facing stereo cameras, however, took somewhat longer to emerge. As Gehrig and Franke (2015) note, ‘the first camera-based driver assistance systems were implemented as prototypes within the European project PROMETHEUS in the early 1990s’. ³ The first appearance of such a system within a production car came with the launch of the Subaru Lancaster ADA (Active Driving Assist) in Japan in 1999 (Wikipedia, 2024), to which we shall return below.

Tracing the origins of these (and other) elements of modern-day ADAS systems enables us to locate important ‘sites where the prospect of “machines-as-agents” is produced’ (Lepage-Richer, 2024: 21). These histories reveal the emergence and ongoing development of automotive phase space; they show also how contemporary automotive phase space is composed of all of these elements. Modern ADAS functions as a combination of them.

Subaru is, of course, just one of many auto makers to incorporate ADAS into their vehicles. Of the 10 most popular car brands in Australia of 2024, all are fitted with some form of driver assistance system. And of these 10 brands, 8 have their own badged names for these systems, producing a confusing array of apparently differentiated products: Ford Driver Assist, Hyundai SmartSense, Isuzu IDAS (Intelligent Driver Assistance System), Mazda i-Activesense, MG Pilot Driver Safety Suite, Mitsubishi MiTEC (Mitsubishi Motors Intuitive Technology), Nissan Intelligent Mobility, and Toyota Safety Sense (the remaining two brands, Kia and GWM do not appear to have proprietary or branded ADAS systems). All of these systems are built around similar functions and technologies, albeit in different individual configurations.

What sets Subaru’s EyeSight ADAS apart from many is its reliance on front-facing cameras, which connect to higher-level communications and control systems. Again, however, Subaru is not the only vehicle manufacturer to employ camera vision technology. Tesla integrates three front-mounted cameras (of a total of eight) in their cars and is well known for its reliance on cameras in the place of more expensive sensors, such as LiDAR. Even major motorcycle manufacturers have or are planning to integrate forward-facing cameras into their products (Purvis, 2023). And then there are systems producers and manufacturers (rather than car companies), like Mobileye, that develop ‘bolt-on’ ADAS technology, including cameras (Hind and Gekker, 2024). In this context, Subaru provides a useful case for our purposes, given its long-standing status as an innovative manufacturer of cars featuring ADAS (in 2022 Subaru passed five million global sales of vehicles fitted with EyeSight; Subaru, 2022), and especially their long-term commitment to the development and use of stereo vision technology.

Development of the Subaru EyeSight system

Like many contemporary production car manufacturers, Subaru incorporates a long list of ‘advanced driver assistance systems’ (ADAS) into their vehicles. In Subaru’s case, these are organized into three ‘preventative safety’ (Subaru, 2023) categories: ‘Vision Assist’, ‘Driver Monitoring System’ (DMS), and ‘EyeSight’.

The first, ‘Vision Assist’, includes the following suite of systems: front view monitor (judges the distance of obstacles in front using the front view cameras and a grille-mounted camera), side view monitor (uses footage from a camera in the passenger side door mirror to show distance from the kerb and other objects), reverse automatic braking (activated if an imminent rear collision is detected), blind spot monitoring (alerts the driver to the presence of other vehicles in their visual blind spots), and rear cross traffic alert (warns of approaching vehicles or pedestrians when reversing). Vision Assist relies on a mix of front, rear, and side mirror mounted camera sensors and rear-located radar sensors.

The second, ‘DMS’ or ‘driver focus distraction mitigation system’, uses infrared camera sensors, located within the vehicle cabin, to monitor driver behaviour. Vision captured through these internally directed cabin sensors is linked to distraction and drowsiness warning systems. The DMS monitors the driver for signs of distraction (a prolonged turning of the head or downward glance) and it ‘tracks the driver’s eye activity to calculate the driver’s level of drowsiness’ (Subaru, 2023). The DMS also permits a driver to configure their preferred settings (air conditioning, seat position, door mirror angles, favourite radio station, etc.) through the use of facial recognition – functionalities that Gilliard and Golumbia (2021) would categorize as forms of ‘luxury surveillance’.

The third, ‘EyeSight’, the focus of this article, employs outward-facing camera sensors that are mounted behind the front windscreen and above the internal rear-view mirror. The visual data captured by these camera sensors is used in a range of driver assistance systems, including pre-collision braking and brake assist, pre-collision throttle management (restricts engine output to reduce available acceleration when a collision is imminent), brake light recognition (detects the sudden activation of brake lights in vehicles ahead), lead vehicle start alert (prompts the driver when traffic ahead begins to move again), lane keep assist (recognizes line markings on both sides of the lane and adjusts the vehicle’s position accordingly), and lane sway and lane departure warnings (monitors the vehicle’s position, warning or adjusting its position if it departs a lane).

In addition to the three main systems, further automated features and functions are also incorporated into contemporary Subaru cars, such as adaptive driving beams (utilizing ‘automated beamforming’ technology, a computer-assisted system that aims the headlights in different directions depending on road curvature and road topography (Shakir, 2022)) and high beam assist (automatically adjusts headlamp range, switching between high beam and low beam, depending on the brightness of detected vehicles and certain road conditions).

Subaru Japan initiated development of a stereo vision system for its cars in 1989 (Subaru, 2022). A decade later, the fruits of this development effort were incorporated into the Subaru Lancaster ADA, with ADA standing for Active Driving Assist. This first-generation ADA system consisted of two charge-coupled device cameras that were mounted on either side of the interior rear-view mirror (Wikipedia, 2024), a configuration that Subaru has continued to use ever since (see Figure 2).

Figure 2.

Subaru’s forward-facing EyeSight dual-camera system.

The ADA featured four safety components: lane departure warning, inter-vehicle distance warning (activated when the ADA detects that the car is fast approaching another car in front), dynamic cruise control, and curve alarm/down shift (the ADA compensates for an anticipated loss of traction or under/oversteer on a curve by down-shifting the transmission automatically to provide engine braking to slow the car in the approach to the corner) (Wikipedia, 2024).

The rebadging of ADA as EyeSight was first announced in Japan in 2008 and subsequently introduced into production cars in Australia and other markets from 2012 (Subaru, 2022). EyeSight Vision Enhanced was announced in 2013 and featured stereo cameras with colour recognition, a 40% increase in viewing angle and distance, lane keep and lane departure assist, and improved object recognition (such as detecting car brake lights and traffic lights) (Subaru, 2013). The most widely available 2024 generation of EyeSight, with dual front-mounted stereo cameras, includes the full suite of ADAS functions and features detailed above: pre-collision braking and brake assist, pre-collision throttle management, brake light recognition, lead vehicle start alert, lane keep assist, and lane sway and lane departure assist.

Innovations and enhancements to the EyeSight system are ongoing. The recently released EyeSight X features the standard two front-mounted stereo cameras, plus an additional two rear-located cameras, for full 360° field of sensing (Olgivie Subaru, 2020); where supporting infrastructure permits it, EyeSight X also utilizes three-dimensional (3D) high-precision maps data and road information for each lane (Olgivie Subaru, 2020). In 2023, an additional front-mounted wide-angle mono camera was added to ‘improve the car’s ability to detect pedestrians and cyclists as part of its autonomous emergency braking software’ (Hansen, 2022). Subaru is also in the process of preparing to integrate geocomputational AI technology (Openshaw and Openshaw, 1997) into EyeSight in support of automatic parking assist and to ‘improve computer recognition in hard-to-see situations, like when road markers are covered in snow’ (Greimel and Okamura, 2022). ⁴

Many of these more recent enhancements are the result of arrangements that Subaru has made with specialized hardware manufacturers. For example, since 2020, the EyeSight system has been produced for Subaru by Swedish supplier Veoneer, a subsidiary of American Swedish firm Autoliv (Veoneer, 2020). In 2020, Subaru struck a deal with Xilinx (then a subsidiary and now a discontinued brand of US-based semiconductor producer AMD) to revamp elements of EyeSight hardware (Nellis, 2020). As a result of these deals, ‘integrated in the system package’ that Veoneer provides to Subaru ‘is a Xilinx Zynq XA processor hosting algorithms developed by Subaru’ (Veoneer, 2020). ⁵

These corporate arrangements are not, however, only about the introduction of new or enhanced features into EyeSight. They also enable greater system efficiencies, which are necessary given the volume and complexity of data being ingested and processed in or close to real time. One significant aspect of Subaru’s deal with Xilinx/AMD, for instance, is that it signals a move away from the use of (Hitachi Astemo produced) application-specific integrated circuits (ASICs) in favour of field programmable gate array (FPGA) chips for particular functions. A key reason for this change is that ‘unlike graphics processing units (GPUs) or ASICs, the circuitry inside an FPGA chip is not hard etched’ (Intel, n.d.), which is to say that it can be reprogrammed as required. ⁶ The reprogrammable and reconfigurable nature of an FPGA makes it cost effective and attractive to Subaru given rapid advances in machine learning and machine vision capabilities. A second and significant benefit for Subaru is that ‘FPGAs can inherently provide low latency as well as deterministic latency [latency with a known duration] for real-time applications’, such as ‘action recognition’ when ‘directly ingesting video into the FPGA’ (Intel, n.d.) – something that is vital in an automotive context when making split-second decisions based on continuous camera sensor data. In 2024, Subaru struck a deal with Xilinx parent company AMD to ‘design circuits for a system-on-chip (SoC) that integrates stereo camera recognition processing’ (and improved object recognition in low-visibility scenarios) with ‘the goal of achieving cutting-edge AI inference performance and ultra-low latency processing at low cost’ (Subaru, 2024b). FPGA chips contribute to what Ludovico Rella (following Donald MacKenzie) refers to as the ‘spatial materialities’ of vision-processing computing systems (Rella, 2024: 20, original emphasis). In the case of EyeSight, FPGAs are tasked with performing a kind of spatial double-duty. They form a crucial part of the spatial materialities of computing that comprise the Subaru platform (Hind et al., 2022), powering the in-vehicle sensors and sense-making devices that enable ADAS to operate. And, as part of this automated ADAS system, they work to produce (transduce) automotive spatial materialities in real time as each vehicle travels along.

Operation of the Subaru EyeSight system: The automation of automotive space

Subaru’s EyeSight utilizes and derives its name from stereo vision technology. A stereo vision system relies on two perspectives to establish depth, just as human binocular vision – what we call eyesight in everyday usage – does. As Andrews et al. (2024: 16) note, ‘every movement of an object involves that object moving into some new relation’ with other, surrounding objects. EyeSight derives information about the distance of other objects based on the relative position of parts of these objects within images obtained from two cameras, which mimic binocular stereopsis, or stereo vision. These cameras are positioned parallel to each other at a fixed distance (a baseline, which aids calibration). The two images are analysed algorithmically to note differences (a shift in pixels) and establish common features (‘correspondence’) of the same point in both left and right images. From this process, known as ‘disparity mapping’, it is possible to create ‘depth maps’ that facilitate object detection and distance-to-object estimation (see Singh, 2022) – what Hind (2022: 75) calls the ‘operationalization of distant capacities’ that inform interactions between sensors (EyeSight) and objects (traffic lights, lane markings, pedestrians, and other road users). The whole process described above, it should be remembered, occurs in real time, as the car is moving.

One way we can productively approach EyeSight’s operation is to regard it as a form of ‘sensing practice’. Nicole Starosielski defines sensing practices (after Jennifer Gabrys) as ‘modes of sensation, not molded by a universal reference point but emerging in relation to technologies and environments’ (Starosielski, 2021: 168). This is an apposite formulation as, despite their name, EyeSight’s stereo cameras ‘“sense” rather than see’ (Hind, 2022: 59). An EyeSight-equipped car makes sense of the world by ‘feeling it out through sensors and cameras, capturing data as it does so, then calculating meaning from it in order to come to a decision about whether an action should be taken’ (Duggan, 2024: 161). The system is designed to perform ‘image operations’ (Mackenzie and Munster, 2019) that capture and convey information for in-car processing of orientation (where the car is relative to lane markings and other cars), perception (what is going on around it), and decision (what it will do next) (McCosker and Wilken, 2020: 96). Through the combination of visual data and algorithmic processing, perception (sensing) is here ‘attuned to action, to the dividing up of movement into points on a trajectory so that they can be acted on’ (Amoore, 2020: 16). In exploring EyeSight, we are less concerned with how vehicle sensor systems learn to see (McCosker and Wilken, 2020) than we are with what they see, what happens to the information when they see, and why this matters.

In Nonhuman Photography, Joanna Zylinska (2017: 51) describes automated vision capture as a form of photography that is not ‘by or for the human’. In the case of ADAS, like EyeSight, this statement needs some qualification. The visual imagery that ADAS cameras capture is not seen by the human driver. So long as it is not manually deactivated by the driver, camera sensor data is automatically recorded and processed by the EyeSight system. Once captured, this data should be regarded as ‘invisual’ in the sense in which Adrian Mackenzie and Anna Munster (2019) mean it, as nonoptical imagery ingested through and operationalized within the automated processes of digital platforms. Visual sensor data persists within the EyeSight platform, but ‘it does not feature as a visual image in the optical or experiential sense’ (Parikka, 2023: 22). The EyeSight system is less concerned with the images it captures than is with the ‘situational awareness and (automated) decision-making’ (p. 21) this imagery permits. In the act of its recording, EyeSight sensor imagery ‘disappear[s] into the machinic apparatus’ (Andrejevic, 2020: 108). ⁷ However, this visual data does ultimately resurface for the human driver in the form of notifications, alerts, and suggested or enforced driving modifications. To get at this twin dynamic – sensory data disappearance and reappearance – we draw from Sam Hind’s distinction between vehicle data and driving data. The former is data that ‘lies under the bonnet, (usually) quietly ensuring the vehicle is operating properly’ (Hind, 2021: 4). This is what Colin Koopman (2019: 1333) would call ‘data that precede communication’. Driving data, on the other hand, is data that ‘is presented to the driver to aid decision-making’ (p. 4); it is ‘vehicle data that is activated in the process of driving a vehicle’ (Hind, 2021: 4).

Vehicle data: How ADAS sensors contribute to the automation of automotive space

Once an EyeSight system is installed, it is connected up to the car’s Controller Area Network (or CANBus) system – the ‘unsung hero’ of in-car operations (Goldstein, 2024). The CAN is the car’s centralized communication system, which permits different ECUs to communicate with one another, highlighting the compositional nature of all vehicle sensor data processing (Hind, 2023). The CAN performs this communication via the ‘Bus’, or ‘binary unit system’, which transfers data between parts of a network with the help of cables. ⁸

Within each Subaru car, the CAN is ‘divided into two separate high-speed circuits, the main CAN and the body CAN’ (Goldstein, 2024). These interact as follows:

Critical systems such as the engine, transmission and ABS are controlled through the main CAN. Less critical electrical systems like power windows, turn signals and HVAC [heating, ventilation, air conditioning] are controlled through the body CAN. The main and body CANs are connected through a gateway, which Subaru calls a Body Integrated Unit. The CAN also interacts with the Local Interconnect Network, which is used to connect sensors. (Goldstein, 2024) (See Figure 3)

Figure 3.

Interactions within Subaru’s CAN (Controller Area Network) system, connecting EyeSight camera vision to vehicle operation.

It is the CAN that permits EyeSight to function. As visual data is captured by EyeSight’s camera sensors, this data is ‘transmitted via the CAN to control systems, including the engine, transmission and Vehicle Dynamics Control modules’ (Goldstein, 2024). Common features of EyeSight, like Adaptive Cruise Control, depend on ‘the constant communication between modules and sensors enabled by the CAN’ (Goldstein, 2024).

To couch these processes within Ash’s language of phase spaces, through the EyeSight system, vision sensors enable specific forms of ‘towardness’, or orientations, to things – in this case, line markings, brake lights, traffic signals, and so on. Through ingestion of visual data about these things, the car’s CANBus system establishes its own forms of protentiality (signals, or more accurately data which is algorithmically processed) that then perturb other internal things (in-car ECUs and the functions they control, such as emergency braking, throttle control, or lane departure warnings). Here we see the whole array of modulations which can be represented in an Ashby phase-space. It is through these operations that EyeSight, in conjunction with the car’s internal computational systems, reconfigures (or transduces) the phase spaces that the car is operating within – that is, the relations of proximity and distance disclosed and responded to by the car as the car’s sensors are ‘perturbed’ (Ash, 2019) by the things around it. Again, it is important to remember that these relations are ‘temporary yet successive’ (Ash, 2019: 59), and in a continual process of renegotiation as the vehicle is moving.

The phase-space emphasizes the system’s variance from a stable state and the regulatory interventions required to achieve stability. In an emergency, interaction between components via the CANBus results in adjustments being made to car behaviour and functioning ‘in milliseconds’ (Goldstein, 2024), such as activation of Pre-Collision Braking and Pre-Collision Throttle Management. These operations are substantially automated, but they also generate notifications for the driver. It is during these operations that vehicle data is surfaced as driving data, and it is to this which we shall now turn.

Driving and driving data: How ADAS data is surfaced for the driver, and how the driver experiences the automation of automotive space

Consider the following scenario: a car driving along a road is rapidly approaching the car in front. The front-mounted pair (or trio) of EyeSight camera sensors determines that the distance to the object (the other car) is too short for the velocity travelled (as Andrews et al. (2024: 16) note, ‘the rate of happening of events is always determined by the speed of the constituent objects involved in creating them’). EyeSight detects the front car’s brake lights, concludes a collision is imminent, and activates the emergency braking, while alerting the driver. This is a not uncommon automotive situation, where vehicle data is surfaced as driving data – or put another way, where vehicle data is converted into a form of ‘sensor-mediated communication’ (Özkul et al., 2023) that is conveyed to the driver.

While vehicle data can be surfaced as driving data in a variety of ways (Hind, 2021: 4), with ADAS-related systems, it tends to be conveyed internally. In the case of EyeSight, these surfacings can take a variety of forms: sounded alarms; visual alerts, which are either displayed on the dash or projected from a top-of-dashboard unit onto the interior of the windscreen in front of the driver; or as tactile feedback. With respect to the last of these, a lane departure is experienced as a kind of subtle nudge that is felt through the hands on the steering wheel as a slight but direct correction to or adjustment of the car’s trajectory. When ABS is activated, tactile feedback is experienced as a violent shuddering of steering wheel, and brake and accelerator pedals. Such notifications, Hind (2021: 10–11) suggests, ‘function like a recommendation engine’ in that they seek to enrol the driver ‘into certain kinds of driverly attention’ and action and reaction; they speak to Mike Featherstone’s (2004: 9) observation that the human ‘senses’ are being ‘reconfigured and extended’ through vehicle technology, requiring us to adopt ‘a more flexible driving habitus’.

In addressing the driver in relation to ADAS technology, it is also valuable to expand consideration of driver data by accounting for the ‘driving experience’ (Hind, 2022: 19). From informal conversations with people about their experiences of ADAS during preparation of this article, a common theme we encountered was irritation with the car correcting their driving behaviour (often aimed, it would seem, at lane departure warnings and adaptive cruise control) and frustration that these systems could not be readily disabled, adjusted by the driver, or disconnected entirely. Peter Merriman (2009: 590) observes how drivers develop certain ‘habitudes’. These are formed through the ‘complex, knowledgeable, embodied ways’ that ‘drivers engage with their vehicles, other vehicles, and the spaces of the road’ (p. 592). Driving habitudes require ‘specific skills, knowledge and forms of spatial awareness’ (p. 592) that are developed over time. Once learned, these driving habitudes become ‘unreflexively embodied’ (Edensor, 2004: 112), constituting ‘deeply entrenched’ ways of inhabiting and interacting with the world (Sheller, 2007: 180). With each new automotive technological development, however, ADAS systems prompt (and at times forcefully introduce) ‘new bodily horizons and orientations’ (Thrift, 2004b: 49), involving ‘significant transformations’ at the level of driver-car-software interactions. Driving adjustments are needed here (‘the capacities and spatialities of the vehicle become incorporated into the embodied sensibilities, capacities and ontology of the driver’; Merriman, 2009: 592), but these adjustments often come reluctantly, as we saw above in the anecdotal accounts of driver frustration. What drivers of ADAS-equipped cars have to come to terms with is the requirement to ‘relinquish control of the “vehicle’s judgement”’ (Sheller, 2007: 187), or, at very least, accept that ‘more of the judgement involved in driving is now being either imposed or managed by software’ (Thrift, 2008: 85). This is because, with ADAS, as Dodge and Kitchin (2007: 267) put it, ‘the calculative power of code [is] supplanting the cognitive ability of the human’. Within ADAS systems, software are ‘active intermediaries’ (Thrift, 2004b: 48). And one of the intended actions of ADAS systems, Sheller (2007: 187) suggests, is to (subtly or less subtly) ‘disengage the work of the driver from the work of driving’.

From our own experiences with EyeSight, there are certain scenarios, however, where the system’s capacity for protentiality – anticipating what comes next – are not always aligned with actual road conditions or a driver’s intentions. Hind describes these moments as ‘loops of disintegration’, where ‘sensor data is poorly captured, objects incorrectly identified, and decisions wrongly executed’ (Hind, 2022: 65). For example, EyeSight sometimes struggles on dual-lane roads where the two lanes merge into one. As the lanes merge, the intervening line markings separating the two lanes disappear, and EyeSight will often misrecognize this as a lane departure and issue a visual and audible warning. And, when approaching a sharp turn too rapidly, such as a hair-pin bend with a high bank or wall in front of the car just as the road turns, EyeSight can have difficulty identifying the curvature or sweep in the road that the car is naturally about to follow, and instead reads the bank or wall looming in front of the car as an object about to be hit, and will activate ABS. Similarly, if you weave in and out of gaps between parked cars staggered along a narrow street too quickly, EyeSight sometimes interprets this manoeuvre as impending impact with an object (a head-on collision with a parked car) and, again, sounds a warning and applies the emergency ABS. Ash (2019) notes that vehicle sensors work to construct an ‘intelligibility of space’ (p. 74). Yet, as he also observes, how humans and smart objects (in this case, cars) experience movement-space are two quite different things (p. 60). This discrepancy is borne out in the above cases, where the ‘“towardness” to other things’ that are ahead and around the vehicle that were detected by EyeSight did not always align with the occupant-driver’s intelligibility and perception of automotive movement-space. Even so, these particular cases do not diminish the overall ways that EyeSight reconfigure the calculation, navigation, and experience of automotive movement-space.

Conclusion

In this article, we have described the emergence of contemporary production cars as dense sensor-computational systems, and how that matters for the experience and governance of automotive mobility. Cars have long relied on sensing devices, but the proliferation of sensors coupled to automated decision-making systems is a recent development, with ADAS now widely integrated into mass-market vehicles. While automotive automation is now often seen in terms of autonomy, with strong public and commercial interest in the radical industrial and economic shifts involved in robotic vehicles, our focus has been on what we now consider low-level automation, where significant automated interventions already occur every day in the work of driving.

We have shown how the contemporary automobile can be understood as a ‘smart object’. We locate this formulation within a well-developed literature in geography, digital media, and related disciplines that understand software-encoded technologies as generating new spatialities, new calculative modes of deliberation, and new forms of machine-readable geography. In ways first theorized over two decades ago by geographers and social scientists, including Thrift, Dodge and Kitchin, Sheller, Urry, and others, ADAS-equipped vehicles now both generate and act upon relations of proximity and distance, and these relations are continuously reformulated as a vehicle moves. We have then used Ash’s more recent concepts of intentionality and protentiality to inform a more detailed account of how driver assistance systems participate in producing intelligibilities and experiences of space. At the same time, we juxtapose Ash’s phase space with Ashby’s much earlier (and, we believe, unconnected) cybernetic account of phase-space to highlight the centrality of systemic states, regulation, stability, and visibility in our understanding of automative automation.

A brief history of driver assistance systems grounds these claims. Radar, cruise control, and camera-based perception each emerged through distinct developmental pathways, before being recombined into contemporary ADAS suites. This historical perspective matters because it highlights the layered composition of automotive ‘phase spaces’: today’s systems are not one technology but a managed assemblage of sensing modalities, control systems, and inherited design decisions. Subaru’s EyeSight is our primary case study, exemplifying a distinctive and long-standing commitment to stereo vision as a primary perceptual system, and illustrating the significance of shifting corporate and technical arrangements. EyeSight’s stereo vision in not ‘seeing’ in any human sense but a sensing practice that generates actionable depth relations through disparity mapping and real-time calculation. The images captured are not for the driver: they are operational inputs, which re-emerge as alerts, nudges, and interventions. EyeSight data travels through the car’s internal communications infrastructure. In Ash’s formulation, the system’s protential orientation to lane markings, brake lights, pedestrians, and other vehicles perturbs control modules at millisecond speeds. EyeSight transduces continuous sensor data into discrete actionable events, where maintaining stability (or preventing collision) becomes a computationally managed regulatory project.

We have argued that the significance of these processes is not confined to internal vehicle data. When EyeSight converts vehicle data into driving data – through warnings, dashboard indicators, heads-up projections, and tactile steering interventions – it reorganizes driver attention and response in ways that resemble recommendation systems, enrolling drivers into specific regimes of perception, anticipation, selection, and correction. This work of enrolment is not always smooth. Drawing on the notion of driving habitudes, the lived experience of ADAS is likely to involve both negotiation and resistance, as drivers adjust to software that increasingly participates in judgement and control. The moments we described – the misreadings of lane merges, sharp bends, and narrow streets – are not merely ‘bugs’ or edge cases. They are revealing instances where the system’s intelligibility of movement-space diverges from the driver’s, demonstrating how automated sensing generates phase spaces that can at times conflict with human judgement.

What follows from this analysis is a reframing of ADAS as a mode of spatial automation rather than simply a suite of safety features. EyeSight does not merely assist the driver within a pre-given space of driving; it participates in producing the conditions under which ‘the road’, ‘the lane’, ‘safe distance’, ‘imminent collision’, and ‘appropriate trajectory’ become actionable objects for intervention. In doing so, it redistributes agency across cameras, processors, in-vehicle networks, actuators, and the driver, recalibrating the thresholds at which the vehicle can legitimately act in advance of a driver, or despite a driver’s action. The automation at stake here is therefore not only the automation of discrete vehicle functions (braking, throttle restriction, steering correction), but also the automation of relations – relations of proximity, distance, speed, and anticipated future states – through which automotive space is rendered calculable and governable in real time.

Both Ash’s and Ashby’s distinct concepts of phase space offer something valuable here: they illuminate different sides of this redistribution of control. Ash’s phase space foregrounds the successive, overlapping, and often non-human relations of proximity and distance that smart objects disclose as they perturb and are perturbed by other objects. It helps us describe how EyeSight creates an automotive space that is not reducible to what a driver sees or intends, and how multiple phase spaces can coexist in the same vehicle at the same moment (human perception, camera-based sensing, haptic intervention, dashboard signalling). Ashby’s phase-space, by contrast, offers a way of thinking about the system at a higher level of abstraction: as a field of possible states and trajectories, and the conditions under which regulation is triggered to return a system towards stability. It makes analytically explicit what is otherwise often submerged in ADAS discourse: the car is continually estimating where it is within a space of states (speed, distance, lane position, object recognition confidence, braking capacity), and ADAS interventions are not merely ‘assistive’, but regulatory actions taken when variance crosses a threshold. Read together, these two phase-space conceptualizations help avoid a common conceptual trap: either treating ADAS as primarily experiential (a driver’s annoyance or reassurance) or treating it as primarily technical (an internal control problem). The point is that these are linked: the experiential surfacing of warnings and nudges is one of the ways the system accomplishes regulation, distributing the work of stabilization across machine components and human attention.

Moreover, Ashby’s parallel formulation of the phase-space enables us to take our analysis a step further: the array of sensors and processing systems which comprise EyeSight not only generate new relations of proximity and distance. As we have seen, the EyeSight system relies on feedback loops to control the vehicle and avoid collisions. A representation of the scope of actual and potential automated control – as Ashby’s phase-space exemplifies – will be essential to our understanding of how the automated system works, and therefore also represents the possibility of external regulatory or governmental oversight. In the simplest terms, this phase-space visualizes what a system can do. We can then rely on the cybernetic principle underlying Ashby’s phase-space: the purpose of a system is what it does, in Stafford Beer’s (2002) famous formulation. ADAS is routinely presented and described as a system designed to safeguard humans from other cars and obstacles on the road (Rose, 2004: 460); in fact, its purpose is to protect the vehicle as far as possible from the unpredictable and potentially catastrophic mistakes that humans make.

ADAS’ transformations of the space of the road involves a set of shifts ‘that asks new questions about the politics and ethics of automated decision-making on the roads’ (Duggan, 2024: 163). While policy and scholarly attention has been mainly focused on smartphones and social media platforms, the modes of governance already embedded in everyday mobility has attracted comparatively limited scrutiny. Nevertheless, EyeSight and its ADAS ilk constitute a pervasive, quietly expanding infrastructure of automated decision-making: one that acts not on feeds and timelines but on trajectories, distances, and embodied practices at speed. If ‘platform governance’ has become a central lens for understanding contemporary power, then the governance of movement-space through in-vehicle sensing and control systems should be treated as an equally consequential domain, with its own forms of opacity, contestability, and harm. This includes not only obvious questions of safety and liability, but also questions of observability, default settings, user override, data ownership and retention, human–machine accountability, and the politics of which forms of perception and driving norms are built into the thresholds that trigger intervention.

Finally, the analytical approach developed here points towards a methodological agenda. Studying ADAS as a producer of phase spaces encourages attention to the interfaces where sensing becomes actionable intervention: the warning chimes, notifications, dashboard icons, haptic steering ‘nudges’, and sudden braking events that punctuate ordinary driving. It also encourages treating misrecognitions and ‘loops of disintegration’ as diagnostic sites rather than anomalies to be dismissed – moments where the system’s phase space is briefly perceptible, and where the gap between human and machinic intelligibilities of space may become observable. Such an agenda can combine close reading of manufacturer documentation, technical analysis of sensing and machine communication, platform data, and ethnographic attention to driver experience and habituation. Taken together, these methods can help make sense of how ADAS systems such as EyeSight create automotive space in real time – how they render it calculable, how they distribute the work of stability, and how they reshape what it means to drive in a world where humans are already sharing the work of judgement with machines.