Autonomous Vehicles: The Human Reality

The promise of autonomous vehicles, gliding silently through city streets, free of human error and the constraints of a driver’s salary, has long captivated the public imagination.

Yet, as Tesla gears up to launch its pilot robotaxi service in Austin, Texas, complete with human “safety drivers” in the passenger seat, it’s a stark reminder that the journey to true autonomy is less a straight line and more a winding road, heavily paved with human assistance.

The “driverless” future, it turns out, is still very much a human-assisted endeavor, grappling with the vast, unpredictable “long tail” of real-world driving scenarios.

Making a vehicle capable of navigating every conceivable road situation with infallible safety remains, for now, a science-fictional aspiration.

No entity is close to cracking that code.

Consequently, every so-called “robocar” currently relies on human intervention in some capacity, from the mundane tasks of cleaning and recharging to critical, life-saving interventions.

The removal of the in-car safety driver is often touted as the “big hard step” – the moment a prototype truly becomes a commercial robotaxi.

But even then, the human element merely shifts, evolving into a complex web of remote oversight and assistance.

Historically, the earliest self-driving prototypes were designed with a human at the wheel, ready to seize control at a moment’s notice.

A quick grab of the steering wheel or a press of the pedals would disengage the autonomous system, reverting to manual control.

Many also sported the ominous “big red button,” an emergency stop that, while rarely used, symbolized the underlying fragility of nascent AI.

For years, all robocar testing outside controlled environments relied on this setup, with teams diligently tracking “interventions” – instances where the human had to take over.

Each significant intervention became a crucial data point, inspiring simulations to understand what the vehicle would have done, and more importantly, what it should have done.

This meticulous process has, by and large, ensured a strong safety record for test fleets.

However, this system is not without its tragic flaws.

The infamous Uber ATG incident, where a distracted safety driver allowed a prototype to strike and kill a pedestrian, remains a grim testament to the fact that even human supervision can fail, especially when protocols are lax or human attention falters.

This incident underscored that the human in the loop, while critical, is also a variable.

As technology advanced, so did the role of the human supervisor.

While Google/Waymo opted for highly trained employee-drivers, Tesla boldly pushed the boundaries by entrusting ordinary, untrained customers with the supervisory role for its Autopilot and Full Self-Driving (FSD) systems.

This move was met with widespread skepticism, yet, surprisingly, it has largely “worked out,” proving to be a safety record roughly comparable to human driving, albeit not the vastly superior performance Tesla misleadingly claims.

Beyond the traditional driver’s seat, the human presence has adapted.

Some vehicles, particularly shuttles designed without conventional controls, place an employee in the passenger seat, armed with an emergency stop button or even a video game controller for manual input.

This approach, seen in early tests by companies like Yandex (now AVRide) and Cruise, serves a dual purpose: it offers a semblance of human control while maintaining the appearance of a “driverless” vehicle – often more for public relations than true safety, given the limitations of passenger-seat intervention compared to a human at the wheel.

Perhaps the most surprising evolution is the advent of remotely driven cars.

Companies like Germany’s Vay utilize this for services like vehicle delivery in Las Vegas.

While remote driving over public networks presents challenges – latency, packet loss, and communication interruptions – systems are designed to revert to a safe stop if connectivity fails.

Economically, it still requires paying a human, eroding some of the cost advantage over traditional ride-hailing.

Yet, it eliminates idle time pay, making it potentially more efficient for fleet operations or specific delivery services.

Waymo, for instance, employs low-speed remote driving for extricating stuck vehicles or navigating complex, software-confounding scenarios, recognizing that at walking speeds, the risk of damage is minimal.

Even lightweight delivery robots, like early Kiwibot and Coco models, were entirely remote-driven, demonstrating the viability of this model in low-risk environments.

The concept of “remote supervision” takes the safety driver out of the car entirely, placing them in a control room, monitoring multiple vehicles via a bank of screens or even VR headsets.

The human supervisor “virtually grabs the wheel” when needed, reacting to problems that are apparent enough to allow for a slight delay in human reaction time.

While no company publicly admits to full-time 1:1 remote supervision, it’s a logical step for pilot programs and a likely silent capability for many autonomous fleets.

These remote monitors typically possess the power to issue a “kill” command, forcing the vehicle into a “minimum risk condition” – a safe stop and pull-over.

The most scalable and widely adopted form of human assistance is “remote assist.” Here, operators in a central command center don’t directly drive the vehicles but offer strategic advice when the AI encounters an unfamiliar or confusing situation.

This might involve suggesting an alternate route, confirming a proposed maneuver, or simply greenlighting the vehicle’s original plan. Companies like Starship Technologies, a delivery robot firm, aim for a 99% autonomy rate, allowing one operator to manage 100 robots – a model that makes human labor cost-effective.

Leaks from Cruise revealed their robots requested help every five minutes during pilot stages, a rate the CEO deemed acceptable for an early phase, with expectations for improvement but never zero.

These remote operators, despite their sophisticated tools, are still human, and human error can occur, as evidenced by a Waymo incident where an operator approved a vehicle’s unsafe maneuver leading to a crash.

Ultimately, the “long tail” of unforeseen circumstances – a sudden road closure, an unexpected obstacle, a nuanced social interaction with a pedestrian – is where human ingenuity continues to reign supreme.

When all else fails, dedicated human rescue teams remain the ultimate fallback, dispatched to manually drive or even tow a stranded vehicle.

For cars without physical controls, these teams carry plug-in game controllers, and in some cases, even law enforcement is equipped with access to such devices.

The journey to true autonomy is not about completely eradicating the human from the equation, but rather redefining their role.

The “driverless” car of the near future will likely remain a deeply human-assisted machine, with operators, supervisors, and troubleshooters working in concert with the AI.

The dream of fully autonomous vehicles may one day be realized, but for now, the ghost in the machine is very much alive, and it’s a human one, sitting comfortably, albeit remotely, at the controls.