The self-driving car is not here yet. Google first began testing its self-driving cars in 2009. The cars, which are now being publicly tested as Waymo cabs in some cities, have driven over seven million miles on American roads since then.
It would take about 300 years for the average American to cover that distance. However impressive that sounds, research indicates it would require about 275 million miles of testing before autonomous vehicles achieve the equivalent of the 2016 US human fatality rate of 1.18 fatalities per 100 million vehicle miles traveled.
Waymo’s autonomous taxicab is, in fact, a tenth or twelfth generation Stanford Cart; the autonomous test vehicle first conceptualized and built at Stanford in 1961. What began as a remote-controlled rover project for one of NASA’s moon missions, soon developed into a platform for autonomous vehicle research.
The eighties and nineties witnessed several degrees of semi and fully autonomous car research from teams at Carnegie Mellon with their NavLab project. In 1995, the NavLab 5, a product of this initiative, was the first car to self-steer across 2800 miles on a trip from Pittsburgh to San Diego.
Yet, it wasn’t until the 2004 DARPA Grand Challenge when self-driving car technology finally took off. The $1 million prize money for the first autonomous vehicle to complete the 142-mile race across the Mojave desert attracted about a hundred registrations, fifteen of whom qualified for the final race. One car covered as much as 7.4 miles, while none completed the course.
By the next year, several teams had incorporated complex LiDAR systems accompanying position and orientation sensors like GPS, accelerometers, and gyroscopes into their vehicles. Unlike conventional cameras and radar sensors, LiDAR (Light Detection And Ranging) sensors allowed the vehicle to map its environment with a high level of granularity and resolution.
The Stanford Stanley that won the 2005 Grand Challenge can be traced back to the original Stanford Cart. When Google finally decided to start its own self-driving project in 2009, it hired from this pool of Grand Challenge veterans. Within the first 18 months, the team had built a system that could navigate around some of the toughest roads including the hairpin-heavy Lombard Street in San Francisco.
Today, there are several car manufacturers that offer driver assists that border on fully-autonomous driving tech, but still, do not qualify as Level 5 automation systems as per standards proposed by the Society of Automotive Engineers (SAE International).
Cadillac’s Super Cruise, Tesla’s Autopilot and Nissan’s ProPilot are consumer-ready, heavily-autonomous driver assistance systems that still offer only limited functionality. They are dependable at lane-keeping and at maintaining a safe distance from the vehicle ahead, however they still require a human at the wheel.
These cars operate on data from ultrasound and radar sensors, neither of which are as effective as LiDAR at discerning smaller obstacles at greater distances.
Laser-shooting LiDAR is still pricey and the chips that process this data are still power-hungry, however companies are working on bringing both of these down. The hardware required for level 5 autonomy is nearly there. It’s the software to correctly interpret all the data from these sensors that is the biggest challenge right now.
The software must broadly perform two distinct tasks: detect and analyze objects from raw sensor data to understand what they are, and mimic human decision-making to prepare the next sequence of maneuvers. Designing a decision-making system is tricky.
A bunch of if-then rules solely cannot cover every possible scenario and implementing a hybrid system, that supplements the if-then rules with an AI engine that can produce smart inferences to take action, will require many million miles of testing and validation to build a library of most possible scenarios.
Apart from this, the software must also include redundancies to ensure there’s a fail-safe mechanism. All of this is not unattainable; however, the on-road testing is time-consuming and requires a large budget for each company to accomplish by themselves. It would be smarter for companies to collaborate and form industry partnerships to this end.
The “driverless-car” revolution is inevitable, though it won’t happen overnight and it most certainly won’t happen tomorrow.
Sources and Further Reading
- The Development of Autonomous Vehicles.
- An Oral History Of The Darpa Grand Challenge, The Grueling Robot Race That Launched The Self-driving Car.