A Future with Driverless Vehicles Requires Sensory Adjustments
Soon after immigrating to the U.S., while taking my first driving exam to get a license, I crashed into a guardrail. But the memory of that nerve-wrecking incident was not the reason I haven’t driven a car for more than 5 miles in the past 23 years. I’m lucky enough to live in San Francisco, a city with a fairly reliable public transit system. I can get around by jumping on a bus, a street car (not named Desire), or a train. So usually I read, fidget with my iPhone, or daydream while I let someone else do the driving.
A few years ago, I discovered that, once the trains go into the tunnel to complete the downtown route, the conductors relinquish control to a computerized system. Essentially, I have been riding in driver-less trains for the past several years now — clearly a prelude to the not-so-distant future where private cars too will become driver-less, or self-driven.
Running a train on a predefined track inside a tunnel, of course, is much simpler than driving a car on a highway shared by others. An autonomous car must have the capacity to make nearly all decisions currently made by a human driver — when to merge, when to get out of the way of another car, when to exit, when to brake, and so on.
The underlying technology in self-driving cars is the specialty of Dr. Sandeep Sovani, manager of global automotive strategy at simulation software maker ANSYS. He’ll be at the upcoming 2012 Automotive Simulation Congress (October 30-31, Detroit, Michigan).
“A driverless car needs to have a good understanding of, and must keep track of, the positions of nearby objects, such as other cars, highway structures, and exits. One of the technologies that enable a car to do that is radar,” said Sovani.
Radar simulation — specifically, simulating the reach and transmission of radar — falls under the electromagnetic segment. ANSYS added this capability to its repertoire when it acquired Ansoft, an electric design automation software maker, for $832 million in 2008.
“You’re actually simulating the electromagnetic waves,” said Sovani.
This includes, for example, the shape of the transmission device and how that influences the way the waves travel in space. To be good enough to replace a human driver who can make instantaneous decisions about navigation, the complex sequence of radar signal processing has to be done in “milliseconds, or less,” said Sovani.
“The radar needs to be designed so it’s emitting the waves in the right direction. Rather than shooting waves into the sky, it needs to be issuing waves relative to the car’s position to find nearly cars,” he added. “You also need to find out how different objects will reflect the waves, so you can decide where exactly to place sensors to capture the returning signals.”
The other important aspect of the self-driving car is its ability to communicate with nearby cars and nearby structures transmitting navigation data. “It’s all about antenna design,” said Sovani.
Since a self-driving car must be able to “see” it surroundings, the car’s vision will comprise, Sovani predicts, radars that detect nearby objects, cameras that capture road conditions, and GPS devices that tell the car its own location at all times.
“Humans’ bio sensors were developed over a long period,” Sovani pointed out. The car’s sensors are still prepubescent, in a manner of speaking.
Simulating a self-driving car is more than simulating radars and onboard devices. It’s about simulation the entire car as an interconnected system — something often referred to as model-based simulation. “To make sure that the system performs well, you have to simulate the system, not just its components,” Sovani cautioned.
The complexity of system simulation often requires far more computing power than what’s generally available in workstations. It suggests high performance comptuing (HPC) clusters will be an inevitable part of such exercises. The simulation in the picture above, for example, shows the electric field distribution and antenna radiation far field pattern at 1 GHz for a complete vehicle simulation according to ISO 11451-2 using the conventional finite element method (FEM) approach. The air region was modeled for the entire room, including the absorber elements on the side walls. 89% of the total number of elements were used to model air. The model was solved using the domain decomposition method (DDM) on a HPC system with 12 nodes in 310 minutes with 75 GB of random access memory (RAM).
Confidentiality clauses in contracts prevent Sovani from naming names, but he said nearly all the leading automotive makers are partnering with ANSYS to explore self-driving vehicles.
“The currently available driver-assistant technologies [such as GPS-based navigation] ultimately aim to remove the driver altogether. They will continue to appear in cars at an accelerated pace from now on,” said Sovani.
But don’t expect a commercial autonomous vehicle to show up at a dealer near you in the next few years. “My personal take — we’re still a very long way from commercial application, probably ten years away,” said Sovani. “As a general rule, the last 20% of the development is solving 80% of the problems. And the remaining problems to solve are enormous. Just think of something like a car suddenly getting a flat tire or a little rock lying on the road that needs to be avoided. Thousands of such scenarios need to be resolved properly.”
Though still far away, a self-driving car will be a welcome change for a social media-addict like me. When the technology is reliable enough to detect and avoid obstructions (for example, the guardrail I crashed into) without my help, I can surf Facebook or write a blog post while riding a car.