In 2019, the autonomous vehicle will start to do to the traditional automobile what the horseless carriage did to the horse a century ago: reduce it to a curiosity operated by a small number of enthusiastic hobbyists. And reshape society in the process.
Why 2019? After several years of successful real-world testing, marred by several highly publicized accidents, fully-autonomous vehicles powered by artificial intelligence (AI) are coming to public roads, starting this year.
Over the last few years, we’ve seen the first wave of AI-enabled features in the form of advanced driver-assistance systems (ADAS). ADAS features, such as adaptive cruise control, lane assist, and blind spot notification are becoming commonplace in otherwise traditional vehicles. The most capable ADAS installations satisfy SAE J3016 Level 3 or Level 4 automation standards, which allow autonomous operation under tightly defined conditions. But these new vehicles will operate autonomously on all roads in all conditions without the possibility of human intervention: no steering wheel and no pedals.
The AI Infrastructure Behind Autonomous Vehicles
Upgrading AI to handle full autonomy will require a combination of cloud-based AI and in-vehicle units.
Deep learning requires two stages: training and inference. During training, the learning model, typically a neural network, learns characteristic features and patterns using a large set of representative data. During inference, the trained model processes a new data set and evaluates it against the existing patterns.
The optimum hardware is not the same for the two operations. The training of the deep-learning model is highly computationally intensive, so it usually occurs in the cloud using high-power servers equipped with graphics processing units (GPUs). Real-time performance or power consumption is not an issue during this phase. However, during inference, when a device executes an algorithm (lane detection, say), real-time performance and power consumption are both important.
Manufacturers have differing approaches to this market.
Some suppliers are going for the whole ball of wax; NVIDIA®, for example, are looking to expand beyond their dominant position in data-center GPUs with automotive-centric platforms such as the DRIVE AGX Pegasus™. Two Xavier™ processors and two TensorCore GPUs form the Pegasus architectural core; the system achieves 320 TOPS (tera operations per second) of deep learning while running an array of deep neural networks simultaneously.
Other companies, such as Texas Instruments, are pursuing a bottom-up strategy. They’re concentrating on the inference piece of the puzzle and leveraging their long involvement in ADAS applications such as laser-based light distance and ranging (lidar) systems. Key to TI’s effort is the TIDL deep learning component suite that includes an optimized set of deep-learning primitives; TIDL enables deep learning on the TDAx ADAS system-on-chip (SoC) products that integrate ARM-based CPU cores, digital signal processors (DSPs), graphics accelerators, and dedicated processing engines for tasks such as vision processing.
Autonomous Vehicles: Where and When
The trickle of autonomous vehicles presently cruising America’s roads will soon become a tsunami. Waymo, formerly part of Google, has announced plans to purchase 62,000 Chrysler Pacifica minivans to add to their 20,000 SUVs from Jaguar Land Rover. To prepare for the rollout, as of October 2018, they’ve completed 10 million miles of driving on public roads and over 7 billion virtual miles under simulated conditions.
And that’s just one company. Toyota and Uber have announced plans to collaborate on developing a self-driving car. And GM’s subsidiary, Cruise Automation, has teamed up with Honda in a 12-year $2.75 billion deal.
Incidentally, AI isn’t just limited to the cockpit: it’s already being used to predict and detect traffic accidents and road conditions. The Surtrac adaptive traffic signal control system, from Carnegie-Mellon University spinoff Rapid Flow Technologies, senses traffic flow at an intersection with the aid of cameras, radar, induction loops, and other inputs. It then creates and implements an optimization plan to route traffic through the intersection and communicates the data to controllers at neighboring intersections.
Further on Down the Road
When autonomous vehicles become commonplace, they will start to change the broader transportation ecosystem. Once transportation as a service (TaaS) becomes the dominant model, it’ll likely spell doom for many transportation-related occupations, beginning with taxi, bus, limo, and truck drivers, plus drivers for USPS, UPS, FedEx, Pizza Hut, etc.
Longer term, the economic case for private ownership of vehicles begins to dissolve. It doesn’t look good for car dealerships, auto manufacturers, and insurance companies. Let’s not forget law enforcement: traffic police, traffic courts, and so on.
On a happier note, widespread adoption of autonomous vehicles will reduce the need for trauma surgeons and nurses. There were 40,000 traffic fatalities in the U.S. alone in 2017. The leading causes of accidents? Human drivers: distracted driving (texting, eating, calling), speeding, and driving while under the influence.
On the legal front, the existing legal framework constitutes one of main barriers to widespread adoption. In the U.S., automakers must satisfy both federal vehicle safety regulations and a patchwork of state laws. Cruise Automation has asked the National Highway Traffic Safety Administration (NTHSA) to approve a car without a steering wheel, brake pedal, accelerator, and other human-operated controls to be used as a robot taxi in 2019; lacking a federal override, many state laws must also be changed.
Intelligent vehicles raise wide-ranging privacy concerns. Surveillance cameras notwithstanding, car trips today are mostly private and anonymous. Your ride-hailed autonomous vehicle will know every detail of your journey and your behavior in the cab. If tough-on-crime advocates have their way, your ride is also likely to deposit you at the nearest police station if needed to answer for past crimes. Thankfully unpaid parking tickets will be a thing of the past.
On a related topic, autonomous vehicles are expected to create a lot of employment for lawyers. Lots of new laws to be written. Lots of deep pockets to go after. Perhaps we’re in the wrong profession.