thought^ Herculano-Houzei’s research demonstrates why robotic brains also need to devote a significant amount of their artificial "neurons” to overseeing the robot’s perception and physical movement. On the day we visited NREC* we walked away from CHIMP’S demo with new insight into robotic operating systems. Given Moravec's paradox and the computational power required to automate simple physical movements, it would not be unreasonable to wonder whether it may simply not be possible to build a robotic operating system that’s fast enough. reliable enough. and smart enough to guide a car, particularly considering the high bar that driverless cars will have to dear when it comes to their safety and reliability. Our next and final stop at KRKC, however, refueled our confidence. Hope appeared in the form of a bright green .John Deere tractor, We were introduced to Carl Wellington, another NREC researcher, whose work focuses on helping agricultural companies automate farming equipment. Modem agricultural practices are technologically sophisticated, using an approach called precision agriculture. Carl explained that farmers have used partially autonomous vehicles for more than a decade. In the early stages of automation, farmers equipped their tractors with highly accurate GPS systems and used farm management software to map out their fields.. In the early days of precision agriculture, a human driver was still required. The first generation of autonomous agricultural vehicles, while capable of handling the straightaways, still needed a human driver to steer the tractor around the curves at the end of each field. A few years later, however, conunercial software capable of guiding an autonomous tractor around curves became available. As autonomous tractors have proven to he safe and capable drivers, the notion of full autonomy has gained acceptance. Farmers are all too happy to have a robot help them with the tedious, yet critical, work of plowing, planting, and spraying fields. “With automated steering on the tractor, the human operator can focus on other tasks," Carl explained. Today companies such as .John Deere and its competitors sell a wide variety of different types of self-driving modules fur their tractors that have proven to be safe and capable drivers. As young people in rural areas migrate away to cities in search of work, self-guided tractors have become standard equipment on many modern farms. Given their punishing schedules, hard-working farmers are all too happy to have a robot help them with the time-consuming, yet critical, work of plowing, planting, and spraying their fieldsAs we wrapped up our interview with Carl, he summed up the appeal of any sort of selfguided equipment in today’s time-starved world; "With automated steering, the human operator can focus on other tasks.'" Controls: The All mash up On the spectrum of engineering challenges;, creating an operating system to guide drivprlcss cans lies somewhere between that of writing code for CHIMP and programming an autonomous tractor. The operating system for a driverless car spans two large and diverse research areas, One is eontnofs eFitjineeruitj, a field of engineering that deals with the regulation of the performance of mechanical parts. The second is the study of artificial intelligence. Controls engineering deals with complex systems (for example, mechanical systems such as robots') that interact in some way with their surrounding environment via inputs and outputs. To impose some order on to a complex system, roboticists organize activities into fnii'-iet'ef controls and high-level controls. In a driver!ess car, low-level controls govern the way the car regulates Its internal systems, such as brakes, acceleration, and steering. High-level controls govern the car's longer-term strategic plans, for example, its navigation and route¬ planning activities. While controls engineering focuses on applying software to manage complex systems, researchers in the closely related held of artificial intelligence strive to build computer software that is capable of intelligent behavior, a broad and vague definition that reflects the field’s staggering breadth and diversity. Al research uses theory borrowed from several other fields, ranging from psychology to linguistics to statistics. Although the quest to create software that exhibits so-called general intelligence is still among the field's long¬ term goals, much of modern Al research homes in on a particular problem space, for example, making industrial processes more efficient, or enab I mg cars to safely guide themselves. An in-depth examination of artificial-intelligence tech niques is beyond the scope of this book. To simplify matters, we divide the rich diversity of Al techniques into two major groups: rule-based, or symbolic Al, and bottom up. or dafa-itriven AJ, also increasingly known as machine learning. Symbolic Al involves breaking down a complex situation or task into atbrmal set of rules that a human programmer writes into software code. In contrast, data-driven AJ (or machine learning) involves the application of algorithms to large amounts of data and uses statistical techniques to classify, rant, or otherwise parse that data. No form of artificial intelligence is innately superior to another. What matters is applying the best At for the particular task at hand. What all AI programs attempt to do is to break the complicated and ultimately unknowable “real world” into a finite number of logical “chunks’1 that can then be processed by software. Fach chunk, or distinct situation, is called a state. A state can be a particular configuration of chess pieces on a board, or it can be a split second in which physical objects are frozen in a particular configuration. The set of all possible situations in a particular problem space is called a stare space, Symbolic AT techniques work best in smaller state spaces, situations in which all possible outcomes can be anticipated and then addressed with formal rules, for example, a factory assembly line has a smaller number of possible state spaces than does a busy city street. Therefore, rule-based artificial intelligence would he an effective technique for the software that guides a factory7 robot through a finite number of possible actions and reactions. Symbolic Al has been the ruling paradigm for decades. At the end of the twentieth century, however, as computing power lias improved and sensors increased the amount of available data from a trickle to a deluge, machine learning emerged from research margins to gain wide acceptance. One of the great advantages of machine-learning is that it doesn't require a human programmer to anticipate every possible outcome of a situation, as is required by traditional symbolic Al techniques. A programmer armed with lots of computation power and a large amount of training data can create a machine-learning software program that “learns" to react to the situation at hand and, in some cases, becomes capable of reacting to novel and unfamiliar situations. A car's operating system is threaded through with different types of artificial-intelligence software that manages all of its various control functions. The autonomous equivalent of modern Homo sapiens. Coogle’s wide-eyed little autonomous buggy, did not burst fully formed from the well-guarded labs of ingenious Google researchers. Instead, Google’s modern robotic car is the beneficiary' of nearly a century of Al and robotics research and contains traces of robotic DNA from several long-extinct research projects, each contributing a key concept here, or a breakthrough technology there. In the popular imagination, there * a long-cherished but inaccurate belief that Homo sapiens, thanks to their innate superiority, managed to remain viable while earlier, more primitive ffomo specie* went extinct. Modern UNA analysis is proving that this notion is flawed. Neanderthals, once thought to have been driven to extinction by their superior cousins,, still lurk among us, UNA analysis reveals that people of European and Asian origin cany hits and pieces of Neanderthal genetic material, suggesting that the process of human evolution is not as staged and linear as once believed. The origin of the software that guides the modern autonomous vehicle is similarly murky’, debatable, and complex. Googled low-level controls, the software that oversees its system of brakes, steering, and speed control, are the intellectual offspring of a primitive military robot, the Dog of War built in 1912. To guide the car along the optimal route, Google's high level control software uses decades-old search algorithms, Some of Google’s car’s ability to learn by comparing current driving scenarios to past experiences evolved from machinelearning techniques that were initially developed in the 1950s. Accelerating, braking, and turning Let's begin with low-level controls. The function of low-level controls is to pull a system back to an optimal set point. Modern feedback control devices are an unheralded but ubiquitous army of referees, constantly adjusting an engine s fuel injection, regulating the voltage in manufacturing machines, even keeping the temperature in your home exactly at the degree to which you set the thermostat. These feedback controls strive to maintain a system's balance: any system—be it mechanical or electrical or biological—needs to be constantly dragged back to a state of equilibrium. Automotive engineers have been using low-level controls in antilock brakes an demise controls since the l q&Os. On a driverless car, low-level controls manage the car's major hardware subsystems, making sure the car drives precisely along a calculated trajectory, and ensuring that its braking and acceleration are smooth, Low-level controls make split-second decisions by running signals to the car’s computer via the controller area network {CAN) bus. if low-level controls are operating correctly, a