Bosch took us for a ride in its level 3 autonomous car
Bosch provided flights to Frankfurt and three nights’ accommodation for this trip to the Bosch Mobility Experience.
BOXBERG, GERMANY—Are autonomous cars like buses? In one way, yes. You wait ages for a ride in one, and then all of a sudden several show up in short succession. In late June, we went for a spin in Jack, Audi’s level 3 autonomous test vehicle. Then, a couple of weeks later in Germany at the Bosch Mobility Experience, we got to sample another such vehicle.
The latter was a modified Tesla Model S sedan, one of a number of research vehicles that the engineering company is using to develop autonomous driving components and systems to sell to car companies. And as you can see in the video above, it’s quite competent at driving itself. A recent study by Bosch suggests that autonomous driving is going to be a big selling point for customers in the future—54 percent of the 6,000 people the company surveyed said it would increase their interest in buying a new vehicle. It’s a challenge that plays to Bosch’s strengths, according to Stephen Stass, senior vice president of Chassis System Control (responsible globally for driver assistance at Bosch):
We are coming from driver assistance—level 1 with our radar sensors already 25 years ago. Nobody thought about autonomous driving back then; if you’d have asked me 10 years ago, I’d have said it was science fiction. But we do it not only from the point of view of sensors and perception but the actuators as well. We pioneered ESP, ABS, and have acquired a steering company, we want to have the full system available. Please imagine failure-operation brake systems. If you have one failure, how do you get the car to a safe stop? So we come from the system point of view to understand the complete system, but we’re happy to sell it as components or as a full system to OEMs who might not have the resources to develop their own autonomous vehicles.
What are all these levels about?
As regular readers of our self-driving car coverage will have picked up on, autonomous cars come on a spectrum, defined by the Society of Automotive Engineers. At level 0 are cars with absolutely no driver assistance at all; think something like a Caterham Seven, or maybe any car built before the advent of anti-lock brake systems (ABS). At level 1, the driver is supported by various systems like the aforementioned ABS or electronic stability control (ESC), but it’s still the human’s job to pay attention and actually control the car.
Level 2 comes next, in which the vehicle is capable of controlling the steering, throttle, and brakes—think adaptive cruise control and lane keeping—but again it’s the human’s job to maintain situational awareness. Even the most advanced semi-autonomous cars on sale today—including Tesla and its Autopilot—are level 2.
Level 3 is where things start to get really interesting. Under defined sets of conditions like being on a divided lane highway or within a well-mapped, geofenced area, the car can do it all, including monitoring its environment for other cars, hazards on the road, and so on. But the human still has to be in the loop as a fall-back in case there’s a system failure or the conditions for autonomous driving come to an end (say, leaving that highway and reentering two-way traffic).
Beyond level 3 are what we think of as completely autonomous cars; you give it your destination and it takes you there with no human input. Level 4 autonomous cars will still only do that in certain use cases—again, think well-mapped geofenced areas—and level 5 is the full “go anywhere, let me just sleep or watch movies and tell me when we’ve arrived” robotic vehicles.
How does it work?
In order for a car to drive itself, it needs to be able to see the world around it. So Bosch has fitted the Model S with an array of sensors—six radar units, six lidar scanners, and a forward-facing stereo video camera. Note that the lidar scanners, like the one that’s mounted to the front of Audi’s upcoming A8, are simpler than the large, roof-mounted 64-channel Velodyne pucks we’ve become used to seeing on autonomous research cars. Those pucks spin the sensors at high speed to create a 360-degree point cloud of the world around the vehicle, but they’re expensive and don’t integrate well into existing automotive design language (unless we’re talking about police cars.)
While the roof-mounted spinning puck or mast creates a superior view of the world around it, it’s overkill for level 3 applications since the use conditions preclude worrying about pedestrians at busy city crosswalks. (A Bosch test vehicle we saw at another facility, meant for level 4 autonomy, had the traditional roof-mounted spinning lidar mast.)
The inputs from all of these sensors are fused together to create a digital representation of the environment—the sides of the road, lane markings, road signs, other vehicles, pedestrians, and so on—and the computer running the show can calculate the free space within which the car can safely travel without hitting anything. Having multiple fused sensors is vital, according to Stass:
You can easily make mistakes if you have only vision control. What do you do in foggy situations? There, radar is much better, but it has challenges with standstill objects. It’s the opposite with cameras. [Which are good at recognizing static objects but less so with moving ones.] In other words, we try to use each advantage of each measuring principle and put it together to something that afterward is reliable. It’s not only about distance [from the car] but having a reliable environmental model. The driving model gets the free space, where you can go without any collisions.
A display on the dash lets the driver know when the vehicle is approaching a stretch of road where it can take over, and, similar to Jack, the Bosch vehicle’s self-driving mode is engaged with the push of a pair of buttons on the steering wheel. The car’s UI changes slightly to signal that it has taken control, and, based on the route, it calculates how long it will be before the driver needs to retake the wheel.
As that countdown gets closer to zero—starting at 30 seconds—the car begins to prompt the driver with a series of evermore insistent visual and audio reminders that it’s time to get ready to drive. A driver-facing camera mounted on the instrument panel uses gaze-tracking to ascertain if the driver is actually paying attention. Should the human not take over, the car turns on its hazard lights and comes to a stop on the side of the road, ready to alert emergency services.
On the demonstration drive, the system performed flawlessly. Vehicle-to-vehicle communication alerted it (and us) to the presence of other vehicles on the road, and the car negotiated the tight turns of the handling track at Bosch’s proving ground without issue.
How do you know it’s safe?
As with some other areas of rapid technological development, it’s not the technology that’s going to limit adoption. Instead, various socio-ethical-legal factors are going to be the rate-limiter. Do the intended users and the regulatory bodies with oversight trust that the new technology is safe? For self-driving cars, that means proving that the systems are at least as safe as human drivers, even if that’s a low bar, considering more than 35,000 people lost their lives on US roads in 2016.
The problem of demonstrating safety sufficiently was highlighted in a study by the RAND Corporation last year. Even though humans are imperfect drivers, we manage to average about 100 million miles for every 1.09 fatalities. If you want to show, with a confidence level of 95 percent, that your self-driving system is at least as good as that, RAND calculated that your test fleet would need to complete 275 million failure-free test miles. Obviously, that would take a very large research fleet or a very long time. Some testing can be done in computer simulations, but Stass told us that Bosch has some other approaches as well:
The solution is threefold. To make virtual reality simulations is one fundamental, and in vision computing we’ve made quick progress. But these simulations are still artificial. You need more. Next is new statistical methods—we call it design-based validation. If some design elements are already proven [i.e. you know that your car can pace traffic and stay within its lane in highway traffic] and you can prove that you develop according to those design elements, you can proceed without a bigger validation. You can maybe just validate the edge cases. Finally, there’s connectivity based validation. This is all about collecting data—to go into series production with a certain level of base functionality which you know is safe and then step by step collect the data of difficult traffic situations and improve your algorithms and use over-the-air software updates. [This will be familiar to the Tesla world.] For sure, it must be secure and safe, but then you learn step by step the infinite world around the car.
Now, if only such systems had been available yesterday when I had to drive from New York City back to Washington, DC…
Listing image by Bosch
Bosch took us for a ride in its level 3 autonomous car