A new computer-controlled tractor leaves the driving to the Global Positioning System.

written by Caroline Seydel

illustrated by Erin Herzog

    A farmer sits on his front porch, reading The Economist. His wife is checking her stocks on the Internet. The farmhand is driving to town for chicken feed.

   Who’s plowing the field? The tractor, of course.

   This scenario isn’t as fanciful as it seems. Life on the farm could be changed dramatically by a new kind of tractor – one that plows perfect rows, day or night, all by itself. Using satellite signals for guidance, this new tractor could help ease the monotony of endless plowing.

   Those of us who think corn comes from the grocery store may not appreciate the difficulties farmers face. Driving a tractor isn’t like driving a car. There are no yellow lines. To plow straight, the farmer must use his hood ornament – a gun sight – to line up the tractor with a distant landmark, like a notch in the mountains. By aiming for that notch, he can keep the tractor’s path straight within about a foot.

But he’s also pulling a plow, typically 30 feet wide. It doesn’t matter how straight the tractor drives if the plow jostles off course. The driver must not only stay focused on the distant mountain, but constantly turn and look to make sure the plow (or any other implement he tows) is still in line.

   Poor visibility also makes driving tough. A farmer needs a windless day to spray fertilizer or pesticide. But in areas like California’s Central Valley, a calm day may bring dense fog , making it impossible to visually line up rows.  In the future, an automated tractor could solve these problems. By calculating its current position and the path it needs to follow, a robotic tractor can drive a precise path, perfectly matching rows to one another in the thickest fog or the dark of night.

   In the Global Positioning System (GPS) lab at Stanford University, engineers led by professor Bradford Parkinson have developed an automatic system that uses satellite signals for steering control. Such control, according to the researchers, saves time, increases farm efficiency, and lowers overall costs.

   A GPS receiver relies on a system of 24 satellites Earth. To determine its precise position, it uses a geometric principle called triangulation. The receiver measures its distance from three different satellites, obtaining a unique set of coordinates. Imagine long strings stretching from the receiver to each satellite. The receiver determines the length of each string – the distance to each satellite. Because it knows the positions of the satellites relative to Earth, it can determine the one location on Earth where the ends of all three strings could meet, and that is its location.

   A standard GPS receiver is accurate to within 100 yards. To achieve higher accuracy, the tractor uses a special technique to refine the satellite signals and measure distances more accurately, allowing it to fix its position within one inch. The tractor also needs to know in which direction it’s traveling and how fast. For these measurements, it compares the position of a receiver aboard the tractor with the position of a nearby receiver, called a base station, placed in the field. Comparing the two positions tells the tractor where it is relative to the base station.

   The tractor project evolved out of a system Parkinson’s lab developed for landing planes automatically using GPS. In 1993, when graduate student Mike O’Connor was looking for a project, he and Parkinson thought applying the technology to ground vehicles was an obvious move.  "In manufacturing, anything that’s dull and boring and dangerous is automated," Parkinson says, which naturally brought to mind farm tractors. O’Connor took on the project, and after successfully rigging a golf cart to drive straight lines, he and Parkinson talked John Deere into supplying them with a tractor.

   Bob Mayfield of Deere and Co. says he’s enjoyed working with the Stanford team on the project. "Deere brings the experience of dealing with customers," he says. "We’ve been able to direct them toward the things they need to look out for," such as soil conditions and hillside plowing.  Although Mayfield says that helping Stanford develop a GPS-guided tractor has been worthwhile in "just proving that it was possible", Deere may someday benefit more tangibly from the relationship. If Stanford receives a patent for any of this work, Deere can negotiate exclusive license to that patent. So far, however, no patents have been issued, partly because the basic techniques aren’t new. Triangulation, for instance, has been around for centuries.

   For farmers, the potential benefits of a fully automatic tractor are many. When plowing a field normally, a farmer uses marker arms that stick out from either side of the tractor, marking the edge of the row by dragging a line in the soil. The farmer uses the line to orient himself on the next pass. Each arm costs about $4000, O’Connor says, but the autotractor doesn’t need them.  Consider the increased productivity from day-and-night operation, and farmers could save real money. A tractor with a preplanned path drives faster than a human driver trying to stay on course while monitoring an unwieldy implement. And, unlike a human-driven tractor, the autotractor can retrace its steps flawlessly.

   When plowing beds for row crops, such as corn, the tractor digs beds, where the seeds are planted, and furrows, which are the narrow trenches between the beds. A slight wiggle in the beds results in seeds falling into furrows when the farmer comes back to plant. Seeds that fall on furrows don’t get watered and fertilized efficiently. They don’t grow. For every seed that doesn’t become an ear of corn, the farmer loses money. With an autotractor, the seeds go in the beds every time.

   Early in development, the researchers tested a tractor prototype in a dusty Fresno field. The automated tractor was pitted against an "expert" human driver in a contest of speed and accuracy. The autotractor was accurate to within three or four inches, but the human driver could stay only within six or seven inches, even with good visibility. The average farmer is accurate to within about a foot.

   Most telling, however, were variations in the driver’s speed over time. "It was a revelation," says Parkinson. Several times during the trial, the human driver came to a complete stop. When the researchers asked the driver why he had stopped, they found that dust clouds had become so dense he could not see the front end of the tractor.  "I was astounded," says Parkinson, "because I thought people were exaggerating [about the dust]. They were not exaggerating."

   Parkinson envisions a day when a farmer will electronically monitor a fleet of autotractors. For that to happen, however, the system of GPS receivers and computer processors that makes the autotractor independent must stand up better to the physical challenges of the field.

   The current crop of students in Parkinson’s lab is making the tractor smarter and faster. Graduate student Dave Bevly is adding sensors to fill in if a nearby tree or barn blocks the GPS transmission. Sensors that measure orientation and acceleration allow the tractor to maintain control for 10 to 20 seconds, until the link to the base station is reestablished. So far, his sensors keep the tractor true to about a foot. That’s not accurate enough for plowing row crops, but it’s acceptable for less sensitive tasks. "It’s about as accurate as a human," Bevly says.

   Bevly is also studying how to keep the tractor accurate at heart-pounding speeds of up to 15 miles per hour, up from the current limit of 2 to 6 mph. He’s using mathematical models of the tractor’s physical properties to predict how it will move in a given situation.

Despite all the gadgets, though, the autotractor is just a mindless machine, doing what it’s told. That’s where graduate student Andy Rekow comes in. He’s teaching the tractor how to adapt to its environment.

   To steer the vehicle accurately, the control mechanism must predict how the tractor will react to a command. For instance, a tractor can’t turn on a dime. Turning the wheels 30 degrees may guide the tractor into a gentle U-turn, or it may send the tractor fishtailing out of control, depending on soil conditions, slope, and speed, among other factors.

   Rekow compares the tractor’s task with a human learning to ride a bicycle. The inexperienced rider twists the handlebars hard to the left or right to turn the bike. The rider predicts what the bike will do when he turns the wheel and mentally compares the reaction of the bike to what he expected. If the bike veers off course, he knows to turn the handlebars more gently next time. Similarly, the tractor’s controller must learn how hard to turn the wheels at a given speed to move the tractor a certain way.

   Besides speed, many other factors influence how the vehicle responds to the controller. Rekow has shown that the tractor is up to 25% more accurate if it adapts to variable physical conditions, such as wet soil, rather than simply using the same commands for every situation.

Think of the bicyclist now riding through sand. The wheels won’t behave as they do on pavement. The rider must adjust for different forces acting on the tires. Rekow’s controller must learn to make these adjustments. It compares the actual result of a steering command with the predicted result and issues future commands accordingly. In mud, for instance, the tractor’s front wheels slip more than the controller expects. The controller predicts a certain reaction from the tractor, and one-fifth of a second later, it checks its prediction against the positions of the tractor body, the wheels, and the implement being towed.