Self-Driving Cars

Taught by Professor John Long, Ph.D.

Self-driving cars are one of the most transformative emerging technologies on the planet.

Many of us already have elements of robotic autonomy built into our cars, and we have for years. Creating robotic cars involves incredible challenges. Chief among them is the trade- off of speed versus safety: To drive safely, you slow down—but if you slow down, you take longer to reach your destination. The goal of the modern automotive industry is to enhance both speed and safety at the same time. Nearly every manufacturer has already embedded robotic elements into your car to assist you.

Cruise Control and Similar Features

  • Cruise control is automatic speed control, and it is a robotic system because it involves sensor-guided movements. You let your car take over the gas pedal, allowing it to adjust the throttle to keep your speed steady as you travel up and down hills.
  • Cruise control is an example of a system that works by negative feedback control. A sensor—in this case, the speedometer—is feeding information about the real world back into the system. The controller compares the actual speed to the desired speed. The difference between the actual and desired setting is called an error. Any system that uses negative feedback control is self-regulating and embodies the kind of self-control we see in autonomous robots.
  • The automobile industry has been adding, one by one, different autonomous systems to vehicles. In response to cruise control’s lack of braking, in 1995, Mitsubishi introduced a laser-based adaptive cruise control. The idea behind adaptive cruise control is called object avoidance in robotics: Don’t hit the car in front of you.
  • Most adaptive cruise control systems use radar as their primary sensor. Radar is an active sensor system; an emitter sends out radio waves at a particular frequency, and then a receiver collects the echo, the return of that signal. Using radar, you can measure the distance from you to the nearest object. You can also tell the relative speed of that object, whether it’s heading away from you or toward you.
  • With adaptive cruise control, you set a target speed, like old- fashioned cruise control, but you won’t hit the vehicle in front of you because of the object-avoidance function. You end up following at a set distance behind the vehicle in front of you. That distance is adjusted based on the speed.
  • To avoid collisions, you need to add an actuator to your toolkit. In addition to throttle control, you also need control of the brakes. You’ll often find emergency braking sold as a feature on cars, and it is a part of the whole object-avoidance system.
  • Even though adaptive cruise control is meant for the high-speed movements of freeway driving, some companies, such as Bosch, have a related system called Stop & Go that is built for slow speeds. In heavy traffic, Stop & Go will bring the car to a complete standstill if needed and then initiate movement when the car in front moves. This makes traffic jams far less frustrating if you are behind the wheel.
  • One common cause of accidents is when you change lanes and don’t see another car in your blind spot. Bosch offers Side View Assist. Using four ultrasonic sensors, two on each side of the car, the Side View Assist signals the driver when an object is detected in the blind spot.
  • In cases where the car is simply drifting out of the lane—such as when the driver is sleepy—some companies offer Lane Departure Warning to signal the driver and Lane Keeping Assist to actually take over the steering. Video is a common sensor for lane keeping, because most paved roads have clear lines that mark the edges of the road.
  • Park Assist can sense other cars and be used when you want the car to park itself. The same sensors that are used for Side View Assist can help detect the open spot, and then commands are issued to the steering and gas to maneuver the car into position.
  • Because twice the number of accidents happen at night compared to the daytime, one of the most exciting driver assist features is enhanced Night Vision. This system works by combining infrared light and video analysis. Infrared emitters up front send out signals that are read by the infrared video camera. A special screen on the dashboard shows the scene ahead. Using software that recognizes pedestrians, the Night Vision system can also brightly illuminate people so that they can be seen and avoided.
  • With the exception of old-fashioned cruise control, all of these behaviors rely on vision—or a sensory capability that functions like vision, such as radar or infrared imagining. Engineer Ernst Dickmanns was among the first to build a robust, working vision system for robotic cars. In 1986, Dickmanns and his team built a fully autonomous robot car that drove in tests on empty streets in Germany. With participation from the automobile industry, by 1995 vehicles equipped with his dynamic vision system were traveling safely for 1,000 miles at speeds up to 100 miles per hour on public highways.

The DARPA Grand Challenge                     

  • Combining navigation with autonomous driving in robotic cars was the primary challenge put forward in 2004 by the Defense Advanced Research Projects Agency (DARPA) of the U.S. Department of Defense. They ran a robot competition called the DARPA Grand Challenge, which was for fully autonomous robotic ground vehicles. To succeed, a robot had to travel 150 miles, off and on the road, between Las Vegas and Los Angeles.
  • What made the DARPA Grand Challenge more difficult than the work already done by Dickmanns’s team was all of the off-road driving. There were fewer regularities: no signs and no cars to 
 Off-road driving is less structured than on-road driving. Even though 15 robots started, not a single robotic vehicle finished the 150-mile course.
  • Some vehicles were able to navigate to the GPS waypoints, but those robots tended to do a bad job of detecting objects along the path. Other vehicles were better at sensing objects, but they weren’t good at navigating to the waypoints. It was clear that a short-range system like Dickmanns’s dynamic vision needed to be combined with a longer-range navigation system.
  • DARPA decided to hold a similar challenge the next year, in 2005. Given all that had been learned in 2004, this turned out to be a brilliant decision. In 2005, 23 robots started the 132-mile off-road course in the desert of Nevada, and 5 autonomous robotic vehicles finished the course. Because of this huge and positive turnaround, many consider the DARPA challenge of 2005 to be a watershed moment in robotics.
  • Perhaps the most revealing result in 2005 was that the winning robot, Stanley, a modi ed Volkswagen Touareg, hadn’t even competed the year before. Stanley was the brainchild of the Stanford Racing Team, which was led by Sebastian Thrun, then the director of the Stanford Artificial Intelligence Laboratory.
  • Stanley’s team’s approach in a nutshell was to treat autonomous navigation as a software problem. The strategy was that all of the failures of 2004 could be solved by building a better control architecture. The design of a three-module controller, combined with estimating uncertainty and machine learning, were the keys to the game that allowed Stanley to finish the 132-mile off-road course and win the race.
  • Thrun and many members of the Stanford Racing Team took what they had learned with Stanley to Google to build the Google driverless car.

Driverless Cars

  • Nearly all automobile manufacturers have worked on autonomous, driverless cars. They have been helped by components manufacturers, such as Bosch, that are making robotic systems that can be deployed on any vehicles. Like Google’s driverless system, Bosch’s system uses LIDAR, video and radar, to create a dynamic map of the world.
  • Part of the technology transferred from Stanley to the Google driverless car was its drivability map. Google’s testing engineers would ride in the passenger seat, essentially looking at the Google car’s version of Stanley’s drivability map.
  • One of the benefits of driving on roads, as opposed to off-road driving, is that roads offer more regularities in the world, and more structure, from lines to signs and curbs. Google also has mapped roads, so the robotic car doesn’t have to start from scratch the way Stanley had to in the desert.
  • What makes the world of roads more unstructured and, therefore, more challenging is that in suburban and urban settings, roads can be packed with irregular traffic—not only cars and trucks of all sizes, stopping and starting at unexpected times and places, but also dogs, pedestrians, skateboarders, and bicyclists. And their positions are constantly changing.
  • States in the United States began to offer driverless cars the right to operate in 2012, and testing began in the United Kingdom, Singapore, and other countries. One of the great things about these vehicular robots is that this technology can be applied to trains, trucks, and buses as well. So, the potential is to completely overhaul our transportation networks that carry people and goods.

Benefits of Driverless Cars

  • For all vehicles, an immediate benefit that we see, even with driver- assist functions in a semiautonomous vehicle, is safety. While the number of deaths from automobile accidents continues to trend downward as we’ve added safety features such as seat belts, air bags, antilock brakes, and adaptive cruise control, tens of thousands of people still die every year from vehicle-related accidents, and millions more are injured.
  • The promise of driverless cars, trucks, and buses with even more automation is that they would reduce those deaths and injuries much further. But what about cases where some sort of collision remains unavoidable? For those cases, rules about how to have a collision can also be programmed into the world model of the vehicle.
  • In addition to safety, driverless cars could increase the independence and mobility of people who, for a variety of reasons, are unable to drive a car themselves. For example, as we age, what often makes assisted living imperative is reduced mobility—not being able to get to the grocery store or a doctor’s appointment.
  • One of the unexpected consequences of having driverless cars is that we will be able to put many more cars on the road. By some estimates, only five percent of a crowded road is occupied by vehicles. Lanes are much wider than the width of the vehicles, and following distances are kept larger than physically required to compensate for the slow reaction times of human drivers.
  • By using robotic sensors and communication to pack moving cars together more efficiently, we could eventually double or triple the capacity of our existing roads. As the human population expands, this would save countries billions of dollars on unneeded road expansion and would keep land available for farming, housing, wildlife, and other purposes.
  • But with robotic vehicles, the opportunities go far beyond traffic lights. Stop-and-go traffic of all kinds is the worst for mileage, for two reasons: If you are stopped, then you aren’t going anywhere, and when you accelerate, you use more gas than you do when you can simply cruise at constant velocity.
  • If the transportation network is tracking all vehicles on the roads, and pedestrian traffic, that information could be used to calculate and send electronic signals to each vehicle about optimum speed and position, to minimize the time that the traffic on average is stopped. Congestion levels that currently bring traffic to a stop could be managed more efficiently so that traffic keeps moving.

 

A Brief History of Driverless Cars

1986 A fully autonomous robot car begins test-drives on empty streets in Germany.

1994  VaMP and VITA-2, driverless cars, operate safely in traffic for more than 600 miles in Germany.

1995  Adaptive cruise control for cars, using lasers, is introduced by Mitsubishi.

2005 Stanley, a robotic car built by a team from Stanford and Volkswagen, wins the DARPA Grand Challenge by being one of five vehicles to autonomously navigate a 132-mile off-road course through the desert.

2007 Boss, a modified Chevy Tahoe built by Carnegie Mellon and General Motors, wins the DARPA Urban Challenge, navigating a 60-mile course while obeying all traffic laws of California, including avoiding pedestrians.

2012 Nevada and Florida become the first U.S. states to permit testing of autonomous vehicles on ordinary roads; Michigan and California follow in 2013.

2015 The DARPA Robotics Challenge focuses on robots for disaster response.

 

Important Terms

machine learning: Computer programs written to make adjustments to their code, with or without direct feedback from a human, in order to improve performance of the code itself or the robot that the code controls.

Stanley: A fully autonomous car that won the 2005 DARPA Grand Challenge, created by Stanford Racing Team.

 

Other Resources

“Cars That Think.” http://spectrum.ieee.org/blog/cars-that-think. This is the best blog on robotic cars, created and curated by IEEE Spectrum, an engineering magazine.

The Great Robot Race. This NOVA program, available to watch online (http://www.pbs.org/wgbh/nova/darpa/program.html), gets you behind the scenes of the teams and under the hoods of the robots at the Grand Challenge of 2005.

 

Questions to Consider

  1. If you want to build the best robotic car possible, which kind of architecture would be better for the controller: behavior-based or model-based?
  2. In 2014, a robotic car competition sponsored by Hyundai in South Korea found that the cars did quite well on a dry, sunny day. But when the road was wet and the weather was partly cloudy, the robotic cars had far more problems, including several missed turns and other maneuvers that they had handled easily during the nice weather of the previous day. What kind of sensors would you suggest be added or improved to address the problems caused by bad weather conditions?
From the lecture series Robotics
Taught by Professor John Long, Ph.D.

Be the first to comment

Leave a Reply

Your email address will not be published.