How a flock of reclusive seabirds became pioneers of pervasive computing. A case study from the sensor net frontier.
John Anderson is stretched out on his side, his right arm plunged to the shoulder in a narrow hole in the ground. He's grubby, his unwashed hair stuck to his forehead and sandaled feet in dire need of a scrub. His belly sags sideways in the brush. With a hard stare and deep concentration, he manages a few groping words: "Uh there nope." The accent is New Zealand laced with high Brit. As he lifts himself up, pulling his arm – now caked with dirt – out of the stiff foliage, he laments, "If we could just understand why here, why this place, it would be invaluable."
This place is Great Duck Island, a 220-acre arc of land off the coast of Maine with no year-round inhabitants and a summer population numbering in the single digits. It's served by solar panels that light up a handful of buildings, crude roads navigable only by tractor, and a boat that floats you to Bar Harbor in an hour and a half. There's no running water. On a clear day, you can see Mount Desert Rock, the most remote lighthouse on the East Coast.
| How It Works: 
Peter Scott On Maine's Great Duck Island, biologists put sensor devices in the underground nests of the storm petrel (1) and on 4-inch stilts placed just outside their burrows (2). These devices record data about the birds and relay it, bucket-brigade style, to a gateway node (3), which transmits the info to a laptop in the research station (4), then to a satellite dish (5) and, ultimately, to a lab in California.
I've come here to watch ornithologist Anderson and three Berkeley computer engineers install an early version of a wireless sensor network. This technology, made possible by the shrinking of the microchip and advances in radio science, is the next step toward pervasive computing. Cheap, mobile, and highly scalable, it's the best hope for dispersing information in a range of environments – office towers, vineyards, hospitals, caves, kitchens, battlefields, even the nesting grounds of birds.
It's those nesting grounds, rugged and isolated, that make Great Duck an ideal test bed for such a system. The Great Duck project will help determine whether this white-board concept – a battery-powered sensor network for collecting scientific data – can actually work in practice. The undertaking, a joint effort by the Maine-based College of the Atlantic, UC Berkeley, and Intel, aims to monitor the habitat of the Leach's storm petrel, a seabird whose lifestyle, including its preference for nesting in arm-length burrows, has made it nearly impossible to study. The tech team so far has deployed 190 devices, each the size of a shotglass, some inside the petrels' burrows and others just outside the entrances. The little instruments, called nodes or motes, house tiny sensors that monitor barometric pressure, humidity, solar radiation, and temperature. (By watching for temperature spikes inside a burrow, researchers can determine when a petrel is present.) The motes report the readings to a gateway node, sometimes passing the data among themselves, � la bucket brigade, bridging distances of up to 1,000 feet.
"We just don't know," says Anderson, dusting off. This is a sentence he will say over and over during my three days on the island, always emphasizing the last word with a peculiar combination of frustration and delight. And the subtext is always the same: At last, sensor nets will shed light on these most mysterious seabirds. Anderson says the new technology will change biology forever – just as it's likely to change high-end agriculture and civil engineering. "Up until now, a biologist from the 1920s could have dropped into today's world and understood everything we do." He shakes his head. "No longer." The instruments that unobtrusively observe the petrels will unleash a stream of information biologists have craved for decades. When I ask what other tool has delivered a comparable advance in his field, Anderson's answer is succinct and telling: "Binoculars."
Take a sensor, any sensor. The one embedded in the seat of your car, for example, which determines that you're present and that your seat belt should therefore be fastened and your air bag on alert. This sensor performs the same function in the same way on every outing. Powered along with the car's other electrical devices by a battery that's regularly charged, it gathers and dispenses information that travels no farther than a few feet. That's easy.
Now, what if the sensor wasn't stationary and had to transmit information over great distances; if it was required to handle multiple tasks, had no nearby power source, and wasn't readily accessible for repairs? All this would pose significant technical challenges. Nevertheless, researchers in recent years have faced down these constraints to make sensor networks a reality.
An Intel R&D team at UC Berkeley has led the way. Directed by David Culler, the hybrid "lablet" is on the verge of overcoming two great obstacles. First, communications. Scattered in large numbers out of easy reach, environmental sensors must work together, pooling skimpy radio-transmission resources and maintaining the network without human intervention. The solution: ad hoc, self-organizing, multi-hop networking, whereby each small instrument has the capacity to find and then pass messages to its neighbors.
Individual rules programmed into the mote's tiny computer orchestrate the sharing. Like members of a football team, the devices do individual tasks but can rely on other players for help. For instance, each mote might be assigned to record information from its onboard thermometer at a given interval, and then broadcast it. If a certain rank-and-file mote is far from the gateway node, it will locate the best messenger to pass its data along. The distant mote does this by checking the position and health of fellows in the network – information that each device announces regularly. Then, just as a quarterback looks for the most open receiver before making a pass, the mote considers its options. If one neighbor indicates that its last message took four hops to reach the node in charge, and another broadcasts that its last message took only two, our distant reporter chooses the latter.
The second problem, and the one that underlies all the rest, is the matter of fuel. Thousands of motes situated, say, at the tops of trees in an inexplicably ailing forest hardly invite frequent battery replacements. Nor would an extension cord up each trunk be an elegant solution. The trick is to use a small, inconspicuous battery so cleverly, with such thrift, that it can last as long as it's needed.
There are several ways to economize: Keep computations to a minimum; be stingy with the number of readings; compress or constrain the quantity of data sent, and use hops for long distances; and put devices to sleep between duties. Sleeping, not surprisingly, provides the best power savings, so the motes in sensor networks spend 99 percent of the time at rest. This raises another problem: How do you get a sleeper mote to wake up on schedule many times a day? One way is to incorporate a global alarm clock into the system, nudging the nappers when it's time to report new data. But all nodes can't call in at once – that would cause transmission bottlenecks. Meanwhile, some nodes will need to be roused not to do their own chores but to help pass buckets; how to schedule those interruptions?
| 
Michael Schmelling Biologist John Anderson goes elbow-deep in a burrow on Great Duck Island.
"It's a very, very hard computer science problem," says Alan Mainwaring, who works under Culler at the Intel lab and has spent the past two summers on Great Duck. "Should everyone wake up all at once? Should everyone know the whole topology of the network?" Mainwaring and his two colleagues on the project have dispensed with the idea of a universal clock. Their system uses what Culler calls low-power listening, in which motes sleep nearly all the time – but in intervals of milliseconds. This way, neighbors are constantly available for hops, as long as a sender catches another's attention within a minuscule, though extremely frequent, wakeful period. To make sure listeners can't sleep through an important missive, the system affixes a preamble to each message that's longer than the mininaps. When a mote wakes up, the preamble will still be transmitting, signaling the listener to stay alert for an upcoming message.
Once you've got the motes awake, of course, you still can't demand much of them. Every computation, every transmitted byte, has a price in power. Culler's group has dealt with these limitations by creating TinyOS, an extremely simple open source operating system. This code manages the radio functions of the machines and handles data drawn from the sensors (converting barometer readings, for instance, from analog to digital, then storing them, compressing them, or just passing them on). It allows motes to locate neighbors, assembles messages, and determines routes. All of this with the simplest, lightest logic system. An entire TinyOS message requires about as much space as the routing instructions alone for a standard email.
Out on the island, TinyOS is gathering data from seven different types of jelly bean-sized sensors. Some are installed on motes planted inside petrel burrows. Some stand aboveground on 4-inch-tall wire posts, recording nearby conditions. Every 5 minutes, each mote sends its observations to the gateway mote, which has a wide-area antenna and plenty of juice from a set of solar panels. It passes the data to two more powerful directional antennas, also sun-fueled, that dispatch the packets to an even larger antenna sprouting from the weathered research station. Laptops inside the building forward the data again, this time to a satellite dish facing out to sea. At one point last summer, 102 motes were beaming information across 50,000 miles, from the meadows of Great Duck into space and then to the lab in Berkeley.
"Intel had this cool technology, but they didn't have any questions to answer. I have endless questions," Anderson says. We are walking from the sensor patch toward the gabled white home that served as residence for the island's head lighthouse keeper until 1986. Today the stark building contains a half-dozen bare mattresses, a single sunken couch, and a table around which the three Berkeley geeks sit, staring at their laptops. "What climates make the chicks thrive?" Anderson asks, giving me a taste of the endless questions. "Which burrows are preferable? Why do birds drag little spruce pine-cones into their nests? We just don't know."
Anderson's relentless curiosity is well suited to tracking these elusive birds. Called petrels because, like the apostle Peter, they seem to walk on water as they skitter across the surface looking for food, the fluffy, black, 2-ounce creatures live dozens of miles out at sea. Unlike the albatross, they're rarely seen hovering in the wake of boats. When they come to land to roost for seven months each year, they huddle in tunnels all day. Only very late, long after dark, do they emerge, flitting across the sky like bats, heading out to sea to harvest food. Studies of some migratory birds – varieties with legs hefty enough that researchers can ring them with electronic tags – reveal their movements, nesting preferences, yearly itineraries. For creatures as small as petrels, however, there's no good way to track them. Even the simplest analysis is tricky: Unlike other birds that come yearly to Great Duck – roughly 1,000 guillemots, 1,300 eiders, 1,200 herring gulls, and 50 black-backed gulls – it's awfully hard to count petrels. In 80 years of study, biologists have had no recourse but to reach down into their nests and feel around. One College of the Atlantic grad student made a valiant attempt to tally the birds by threading a camera typically used to examine sewer pipes into several of the island's burrows. She estimated there were 9,300 pairs of the birds, give or take 6,500. It's the best reckoning Anderson has ever gotten. "We definitely have the largest known population in the Lower 48," he says. "Maybe there are others. We just don't know."
The sensor net will finally get Anderson up close to his mysterious animals. Already, it has produced data that helps explain the birds' nesting choices. Despite discrepancies in temperature readings from the aboveground motes, the burrows' interior conditions are proving highly consistent. Whether the outer motes are planted in the warm air of the meadow or the cool shadows of the forest, the burrows' inner chambers remain about 54 degrees Fahrenheit. What seems to matter, then, isn't so much the island's microclimates as its soil.
The sensor technology has also opened new avenues of inquiry. Watching the temperature readings climb and fall over the course of days, Anderson has confirmed that petrel parents spend an unusual amount of time away from their eggs during incubation and from their chicks once they're hatched; neither eggs nor chicks, it seems, mind the cold. "This," he says, "raises some important physiological and developmental questions."
In the long run, Anderson hopes to use the sensor network to locate undiscovered groups of petrels, as well as roosting populations of other species on islands where landing conditions permit only a visit or two each year. Finding thousands of petrels living differently elsewhere – tunneling into less spongy soil, for instance, or settling in chillier burrows – would go a long way toward revealing the shy creatures' habits. Instead of checking on the birds one at a time or even monitoring them remotely by the hundreds, Anderson wants to compare and analyze the behavior of thousands of birds over a dozen islands.
There are eight of us wandering about in the woods, the peaty soil sinking several inches under our feet. The gray light of an overcast sky is filtering through ancient mossy growth. Up ahead, Anderson is on his hands and knees, crawling under the lowest branches of a spruce looking for an active burrow. (If a sprinkle of freshly dug soil has collected at the mouth of a burrow, it's likely some petrel has chosen it as home.) Several COA students follow behind, one planting a numbered red flag at every spot where Anderson decides to place a mote; one carrying a GPS receiver, much like a flag-bearer for a marching band; and one dutifully recording the coordinates. The device is encased in plastic that allows radio frequency out but won't let the sun's heat in.
A wunderkind Berkeley engineering student who works with Mainwaring, 23-year-old Joe Polastre, urges us to go deeper into the woods. He wants to push the system to test how it will handle multi-hops. Anderson is more interested in a nearby burrow that he believes contains an egg. After a brief conversation, a mildly irritated Anderson gives in and agrees to choose farther-flung nests. The mosquitoes get worse.
What makes this itchy trek in the woods significant for sensor networks is precisely that it takes place in the woods – real, honest-to-goodness woods. So far, developers of sensor networks have faced their challenges and celebrated their breakthroughs from the comforts of the lab. They've set up PDA-sized devices around their cubes and in hallways and cheered when they saw their routing schemes function as expected. For Mainwaring, though, the Great Duck network is not a test of whether the system can work. He knows it can. It's a test of whether it works under real conditions. Every rain-soaked or silent instrument is key to the project. "With these motes," he says, reducing the experiment to a single sentence, "it all comes down to one question: What happens when they get dirty?"
That's why, as we pick through the spruces, we pass the elements of an additional layer of technology: heavy plastic suitcases sprouting cords. This secondary setup is devoted solely to verifying the first system. It's made up of five cameras buried in the ground above five different nests, deep enough that the lenses poke slightly into the burrows' inner chambers. These infrared devices will deliver hazy images of any animals present to corroborate the motes' occupancy readings. As temporary test instruments, the cameras are powered separately from the wireless net, using cable that doubles as electrical cord and Ethernet – power in, data out. A large server, lodged in a watertight case, sucks electricity through extension cords that run back to the island's main photovoltaics. This validation system comes with its own set of problems. Says Mainwaring: "You've seen the cute little bunnies? Actually they're wild hares, and they've chewed through our Ethernet."
By the time we finish our hike, we've placed 14 more devices in the ground and fixed several cameras. Back inside, the team stands around a laptop watching numbers appear. "Everyone's reporting," says Polastre. The newly installed motes, loaded with chips and sensors, are sending packets from one to the next, then to the gateway node, then via solar-powered antenna to a database on a computer here in the house. Polastre pulls up a record of motes installed two weeks earlier, showing off a graph of occupant temperatures dropping as an adult petrel departs into the night. Mainwaring launches a fuzzy video feed from one of the validation cameras, revealing in real-time the tiny movements of a bird twitching and breathing as it roosts.
As I leave the island, my feet are covered with bites and the grime around my neck is scary. I can't wait to splash my face in the first sink I find.
We trudge to the boathouse, where a spacious metal rowboat that the captain calls a pea pod is cinched, with the help of a diesel-fueled pulley, to the top of a ramp. After the craft is packed with our gear, one of the students releases the pea pod, which slides at a frightening clip some 160 feet down to the water. In this low tech manner, Great Duck islanders, otherwise trapped on land by a ring of lethal rocks (Anderson warns of RTMs – "rocks that move"), can escape the wilderness.
From the bottom of the ramp, we row out to a 35-foot former lobster boat called Indigo. It will deliver us to the mainland. The surf is bumpy. We pitch unpredictably, our narrow paddles a remarkably inefficient means of propulsion, and I think for a minute of the latest round of sensor readings, which have already reached California.
A Field Guide to Remote Sensing One day, intelligent sensor networks will be full-service remote labs capable of interpreting and acting on information independently. But for now, they're the equivalent of superhuman grad students, collecting data from the far reaches of the planet – volcanoes and vineyards, South Pole ice and Southwest desert – and transmitting it back to the classroom for analysis. Here's a sample of projects in planning, testing, or operation. – Dustin Goot
Agriculture Okanagan Valley, British Columbia The researchers: King Family Farms, Intel Research, AgCanada The launch: Spring 2003 The Scoop: A 65-mote grid spread across more than an acre of vines collects temperature data that helps growers determine which grapes to plant and where to irrigate. The network also delivers frost alerts and tracks temperature accumulation, a metric used by vintners to plan harvests. The Promise: High-yield viticulture. Intelligent networks will manage automated irrigation and crop maintenance schemes customized for each vine.
Testing Water Palmdale, California The researchers: Los Angeles County Sanitation District, UCLA, UC Merced, Loyola Marymount University The launch: Winter 2004 The Scoop: Los Angeles County wants to give treated wastewater to farmers for irrigation, but it must ensure that the water's nitrates don't trickle into groundwater at toxic levels. Instead of testing for nitrates by digging sampling wells, which catch problems after the groundwater is already contaminated, the plan is to bury sensors that track the pollutant as it seeps through the soil. When readings begin to get high, farmers know to stop spraying there for a while. The Promise: The end of the EPA as we know it. Sensors buried near dumping sites will alert polluters when they step over the line.
Testing Air Carson, Washington The researchers: AmeriFlux Network, UCLA, Center for Embedded Networked Sensing The launch: Fall 2003 The Scoop: Sure, forest canopies theoretically work as carbon sinks, but how do you gather data from a hundred feet up? Researchers at Wind River Experimental Forest measure carbon influx with bulky sensors hung on cranes. For more accurate figures, they're building the first self-adapting 3-D sensor network. Laptop-sized processing motes will crawl along cables strung between trees and lower weather and CO2 sensors into the canopy on adjustable wires. The Promise: Geography-based pollution control. A factory's emissions allotment will depend on how much carbon nearby forests can absorb.
Fighting Desertification Chihuahuan Desert, New Mexico The researchers: Long Term Ecological Research Network, Jet Propulsion Laboratory The Launch: Summer 2003 The Scoop: At the Sevilleta Field Station, scientists are studying the creosote bush, a shrub that creates a kind of mini greenhouse effect, making it an early sign, and perhaps a cause, of expanding desert. Temperature, moisture, and light sensors are being placed around three creosote bushes – in the leaves and surrounding soil – and connected to a mote at the central server. The data will be compared with readings taken from juniper and mesquite bushes and from open areas. The Promise: The dune stops here. More broadly, a deeper understanding of microclimates.
Fighting Fire Claremont Canyon, California The Researchers: UC Berkeley The Launch: Summer 2004 The Scoop: How hot is a forest fire? The answer is invaluable, though difficult to obtain. Heat gradients are important for learning where a fire is spreading and when adjacent trees will flare up. UC Berkeley's FireBug project consists of golf ball-sized sensor motes that can be dropped ahead of an inferno and send data back out. The research team has gone through fire training, and a test is planned for a controlled burn next summer. Over time, scientists hope to deploy sensors in the wilderness as a way to anticipate hot spots. The Promise: A new age of firefighting. Proactive monitoring with smart dust will prevent forest fires.
Seismology San Gabriel Mountains, California The Researchers: UCLA Department of Earth and Space Sciences, Center for Embedded Networked Sensing The Launch: Spring 2004 The Scoop: Sensor networks don't just work in small spaces. UCLA professor Paul Davis will install a ladder of 50 networked seismometers – 25 on either side of the San Andreas Fault – to analyze the movement of giant tectonic plates. The ruggedized devices, each the size of a briefcase, will record even faint rumbles, allowing Davis to calculate the depth of the fault and pinpoint locations where stress is accumulating. Understanding stress could lead to better forecasting. The Promise: An early warning before the big one.
Space Exploration MacAlpine Hills, Antarctica The Researchers: Antarctic Search for Meteorites, Jet Propulsion Laboratory The Launch: Winter 2002 The Scoop: Space buffs are turning to sensor networks to advance the search for life on Mars. The plan: establish a virtual presence on the Red Planet through sensor pods, which could scope out terrain and temperature and direct rovers to areas that might support life. Last winter, researchers tested a 14-node cluster in Antarctica, partly to assess the equipment's mettle in cold, harsh environments. On future Antarctic trips, scientists hope to test sensors designed for detecting signs of life such as CO2 or methane. The Promise: Astronauts become obsolete. Rovers will set up mote-enabled virtual labs on the cheap.
