1536-1268/05/$20.00 © 2005 IEEE ■ Published by the IEEE CS and IEEE ComSoc PERVASIVEcomputing 37

ENERGY HARVEST ING & CONSERVAT ION

Battery-Free WirelessIdentification andSensing

Collections of tiny, inexpensive wire-less sensor nodes capable of contin-uous, detailed, and unobtrusivemeasurement have attracted muchattention in the past few years.1

Prototypes exist for applications such as earlydetection of factory equipment failure, opti-mization of building energy use, habitat mon-itoring, microclimate monitoring, and moni-toring structural integrity against earthquakes.

Unfortunately, the very prop-erties that make sensor nodesattractive for these applica-tions—low cost, small size,wireless functioning, and timely,long-term radio communicationof data—make powering thema challenge. Particularly chal-lenging is the production of anenergy source that’s tiny and hasa long mean time to replace-ment yet can supply enoughpower for wireless communica-

tion of sensed data. Recent work has focused onbatteries and ambient power scavenging to pro-vide power for sensor networks (see the “PowerSupply Options” sidebar for a more in-depth dis-cussion of these power sources).

The Wireless Identification and Sensing Platform(WISP) project explores a third approach, basedon passive radio-frequency-identification technol-ogy.2 In traditional passive RFID systems, ambi-ent high-power readers interrogate battery-free

devices, called tags, that modulate the interrogat-ing signal to communicate a unique identifier tothe reader. The WISP project aims to augmentRFID tags with sensors so that tags can also sendsensed data to the readers. We call these augmentedtags wisps. Basing wisps on RFID has some imme-diate advantages. RFID tags communicate to ambi-ent readers over distances of up to eight meters.So, a single US$1,500 four-antenna reader cur-rently available off the shelf should cover fourrooms in a home. The tags can be read at nominalrates of up to 2,000 per second. They have the formfactor of mailing labels, cost roughly $.50 each,and can communicate even when obscured bymany materials (conductive and water-rich mate-rials are notable exceptions). Perhaps most impor-tant, business and manufacturing trends supportthe development and adoption of smaller, cheaper,and longer-range tags. Solutions compatible withRFID standards might therefore find quickeracceptance and see faster improvement thanother solutions. The “Other Approaches toReader-Powered Sensors” sidebar discusses otherapproaches to reader-powered sensing.

Many challenges remain, however. Manyapplications proposed for conventional sensornetworks require instrumentation of large, oftenoutdoor areas—clearly beyond the range ofRFID-style readers. Some applications are alsoliberal in the density and size of tags needed: inmany cases, even $1 to $10 solutions might notbe too expensive, and relatively large devicesmight be acceptable. Furthermore, long-range

Wisps, a class of battery-free wireless sensors, scavenge power fromambient radio-frequency-identification readers to communicate senseddata over room-sized areas. A simple usage model enables a variety ofdistributed sensing applications.

Matthai Philipose and Joshua R.SmithIntel Research Seattle

Bing Jiang, Alexander Mamishev,and Sumit RoyUniversity of Washington

Kishore Sundara-RajanIntel Research Seattle andUniversity of Washington

38 PERVASIVEcomputing www.computer.org/pervasive

ENERGY HARVESTING & CONSERVATION

R ecent distributed sensing work has explored two main power

supply options: batteries1–3 and ambient power scavenging.4,5

Battery technology is mature, extensively commercialized, and com-

pletely self-contained. However, given current energy density and

shelf-life trends,5 even for relatively large batteries and conservative

communication schedules, the mean time to replacement is only a

year or two. The problem is aggravated significantly for batteries with

more inconspicuous form factors. For deployments with hundreds of

sensors, this means that a battery will need a replacement every few

days, an unsustainable rate for many applications. Finally, systems

based on battery-powered communication still cost more than US$5

a unit, with no clear commercial or technical route to a solution that

costs less than a dollar.

Ambient-power scavenging,whichharvestsenergy (especially

solarenergy) fromthesurroundings forwireless sensing,can inprin-

ciple supplypower indefinitely.4,5 Implementationshaveprovided

low-duty-cyclecommunicationatvery small sizeunderambient

conditions thatarereasonable formanyapplications.However,chal-

lengesremain: relianceonambientpower constrainsbothwhere

youcanplace the sensors andwhenyoucanuse them.Further-

more,given thesedevices' useofpower-harvestingcomponents

(suchas solar cells) and relatively sophisticated radios, it remains to

be seenwhether thesedevices canbemanufacturedat their target

costof tensofcentseach.Almost as important, the commercial case

for high-volume manufacture of these devices is still being made.

REFERENCES

1. M. Feldmeier and J. Paradiso, “Giveaway Wireless Sensors for Large-Group Interaction,” Extended Abstracts 2004 Conf. Human Factors inComputing Systems (CHI 2004), ACM Press, 2004, pp. 1291–1292.

2. J. Hill et al., “System Architecture Directions for Networked Sensors,”Proc. 9th Int'l Conf. Architectural Support of Programming Languages andOperating Systems (ASPLOS IX), ACM Press, 2000, pp. 93–104.

3. I. Munguia-Tapia et al., “MITes: Wireless Portable Sensors for StudyingBehavior,” Ubicomp 2004 Adjunct Proc., 2004, http://ubicomp.org/ubicomp2004/adjunct/demos/tapia.pdf.

4. J. Paradiso and M. Feldmeier, “A Compact, Wireless, Self-Powered Push-button Controller,” Proc. 3rd Int'l Conf. Ubiquitous Computing (Ubicomp2001), Springer-Verlag, 2001, pp. 299–304.

5. S. Roundy et al., Energy Scavenging for Wireless Sensor Networks, KluwerAcademic Publishers, 2003.

Power Supply Options

R eversibility, ID space, and range distinguish approaches for ac-

quiring and transmitting sensed data using scavenged signals

from ambient readers.

Reversibility refers to whether the tag can be reused after sensing

the phenomenon of interest. In some cases, because the phenome-

non (commonly, exceeding a threshold temperature for perish-

ables) irreversibly changes the tag structure, the tag lasts for only

one measurement.

ID space refers to the number of distinct identifiers available to

tags. To provide globally unique identification of the sensors and

their associated objects, it's crucial to have at least 32 ID bits. The MIT

Media Lab developed a series of devices that use the chipless RFID

(radio frequency identification) approach, in which the individual tags

have distinct resonances that nearby readers can discern. Variations in

magnetic-coupling strength can indicate tag movement.1 More gen-

erally, the use of smart materials can change the tag's resonance

frequency on the basis of ambient conditions, letting tags measure

temperature, force, or humidity. Unfortunately, the number of distin-

guishable resonances available to these chipless approaches isn't suffi-

cient to enable globally unique identification.2

Finally, to the best of our knowledge, all existing sensors that are

based on RFID use near-field coupling and have ranges of a meter

or less. The project most like the WISP (Wireless Identification and

Sensing Platform—see the main article) project is the University of

Auburn's RFID sensor project, which uses (custom) passive RFID sen-

sors coupled with biosensors to detect food spoilage. The project's

most recent publication sketches a large (saucer-sized) inductively

coupled short-range (10 cm) battery-powered tag, with a stated goal

of removing these constraints.3

REFERENCES

1. J. Paradiso, K. Hsiao, and A. Benbasat, “Tangible Music Interfaces UsingPassive Magnetic Tags,” Proc. ACM Conf. Human Factors in ComputingSystems: Special Workshop on New Interfaces for Musical Expression (CHI2001), ACM Press, 2001.

2. R. Fletcher, Low-Cost Electromagnetic Tagging: Design and Implementa-tion, PhD dissertation, MIT, 2001.

3. S. Nambi et al., “Radio Frequency Identification Sensors,” Proc. 7thWorld Multiconf. Systemics, Cybernetics, & Informatics (SCI 2003), Int'lInst. of Informatics and Systemics, 2003, pp. 386–390.

Other Approaches to Reader-Powered Sensors

RFID tags are susceptible to collision (lim-iting the density of readable tags), harvestvery small amounts of power from thereader’s signal (making it unclear whetherwe can power circuitry to acquire andtransmit extra information from sensors),and are quite sensitive to environmentaleffects such as occlusion and presence ofmetal. Finally, it’s unclear how to con-vert an RFID tag/reader system into awisp transmitting sensed data in a man-ner that preserves the RFID tags’ formfactor and is completely compatible withRFID protocols.

Nevertheless, we argue that distrib-uted sensing based on wisps is useful andfeasible. Our usage model for wispsenables a variety of applications relatedto detecting day-to-day human activities.Our working prototype wisp, a 1-bitaccelerometer we call the �-wisp, is fullycompatible with existing RFID proto-cols and in principle enables these appli-cations. Measurements show that the �-wisp’s performance is, for the most part,in the range required by the applications.

Usage modelWe envision a simple usage model for

wisps. Long-range RFID reader antennasare distributed around the area to beobserved so that readers can provide blan-ket coverage of the area. These readersare typically connected directly to a poweroutlet in the wall, but they can be mobile.Next, we attach wisps to points of inter-est—for example, objects, people, andstructures that either are in the space orwill enter it. Finally, we set the readers toscan the space continuously. The wispsrespond to the reader with their identifierand sensed value. The ID lets the sensedvalue be associated with a unique sensor.The application can then act on the sen-sor snapshot of the space as appropriate.

The �-wisps are intended to measurethe acceleration of the objects to whichthey are affixed. When interrogated, �-wisps transmit a 1 along with an ID if

the object is out of its rest configuration.If the object is in the rest configuration,the �-wisp either transmits a 0 and its IDor doesn’t respond at all, depending onthe application.

ApplicationsThe high-density limited-range wisp

usage model is particularly well suitedto at least one rich, important family ofapplications�that is, indoor humanactivity tracking. The WISP project ispart of the System for Human Activity

Recognition and Prediction. SHARP

models human activities probabilisti-cally in terms of the sequence and dura-tion of the objects’ use during thoseactivities. Given a trace of objects used,an inference engine tracks the likeli-hood of various activities in progress.Early results have shown that giventhese observations, the inference enginecan tell with high certainty and in con-siderable detail what physical activitya person is performing.

SHARP relies on getting traces of objectsactively being used in the activities it’stracking. Currently, the project relies onhand-worn RFID readers to report touch(and therefore use) of RFID-taggedobjects. Wisps promise to remove the needfor the wearable device, which is unsuit-able for many applications. Motion (inparticular, a change in orientation) is typ-ically a good indicator that an object is inuse. Tagging objects of interest with the�-wisp will potentially let us detect objectusage unobtrusively and in great detail.

Caregiving for eldersCaring for elders is an emerging chal-

lenge in many societies.3 A central chal-

lenge in elder care is facilitating theirindependent living while ensuring thatthey’re going about their daily liveswith sufficient competence. To monitoran elder’s activities, caregivers mustspend substantial time near the indi-vidual. Unfortunately, given currentsocial and economic trends, neitherfamily nor professional caregivers canmeet the required time commitment,motivating the need for a semiauto-mated solution.

We’re investigating two applications

in this space. The electronic Activities ofDaily Living4 (eADL) form attempts toautomatically fill in a state-mandatedform requiring professional caregiversto rate their elderly clients’ ability to per-form various activities, such as house-work, making snacks, and grooming.Professional caregivers would use thisform as a starting point when visitingelders, rather than spending time at-tempting to recreate the information. Asecond application, the CareNet Display,provides family members with up-to-date information on whether elder rela-tives completed key activities.5 Familymembers can coordinate elder care usinginformation from the display.

To get the trace in either application,we tag a variety of household objectswith �-wisps and place RFID antennasin room corners. Our experience overthe past year has indicated that sensingthe vast majority of ADLs will probablyrequire tagging about a thousand objectsover the entire house (objects can be assmall as a box of tea or a spatula). At lessthan $0.50 a tag, a thousand tags perhouse seems acceptable for multiple-yearelder care support. Our experiments

JANUARY–MARCH 2005 PERVASIVEcomputing 39

Our usage model for wisps enables a variety of

applications related to detecting day-to-day

human activities.

show that even partial deployments areuseful. Typically, fewer than 10 objectsare in active use at any time.

Another key aspect of this kind ofhome-based deployment is the health riskof having an always-on RFID reader inthe space. Current standards require thatusers be more than 0.5 meter from thereader, implying that readers will have tobe securely mounted high on walls or oth-erwise sufficiently removed from users.

Best-known method managementMost organizations have recognized

experts to perform key tasks. The tech-niques the experts use are termed best-known methods. Managing BKMs,which includes capturing them, trainingworkers to use them, and enabling theiruse on the work floor, is a strategicadvantage for organizations. SHARP isparticularly interested in BKMs relatedto manual activities such as operatingmachinery, cleaning, caring for patients,and assembling parts.

We’re investigating two applicationsin this space. In one application, wemonitor novice anesthesiologists to eval-uate how well they perform relative toexperts.6 Currently, onsite instructorsperform these evaluations in an ad hocmanner. The instructors are receptive toaugmenting the evaluation process withdetailed sensors both because the USgovernment is pushing for more quanti-tative grading criteria and because theyfeel that detailed information couldimprove training.

The second application monitorsworkers in a semiconductor fabricationplant to reduce system downtime due tohuman error. In these plants, the ma-chines are so highly optimized for per-

formance that human error is increas-ingly the dominant cause of downtime.

These applications involve tagging var-ious tools, containers, and machines withwisps to track the progress of the anes-thesiologists and plant workers, and posi-tioning readers over designated workareas. Our preliminary studies suggestthat the presence of large metal structuresin constricted areas and requirements forelectromagnetic shielding complicate

wisp deployments in these milieus. Onthe other hand, because the activities aremuch more structured and the perform-ers more cooperative than in the home-sensing case, we can carefully place read-ers and tags. We would likely tag tens tohundreds of potentially small objects(such as screwdrivers or syringes) in eachwork area.

Personal activity-based promptsProactivity is a stated goal of perva-

sive computing. In the proactive usagemodel, the computer need not wait forexplicit interactive commands from theuser. Rather, it acts autonomously onthe basis of implicit triggers associatedwith user activities. The most commonautonomous action studied so far isprompting users. Our particular interestis in activity-based reminders (for exam-ple, your phone might remind you totake your pills when you sit down to ameal). We envision that simply taggingthe set of objects relevant to an activityshould enable reminders for that activity.

Design requirementsOur early experiences with the appli-

cations suggest some constraints on �-wisp design:

• To minimize the density of RFID read-ers (which serve as wisp readers), wispsshould have a read-range comparableto that of the underlying RFID system.

• Unobtrusively tagging mundane ob-jects requires very small tags. Weshould be able to tag a spatula with lit-tle change in its usability, for example.

• Because we might tag hundreds ofobjects, tags should be very cheap. Adollar is (barely) acceptable for health-care applications; $10 isn’t.

• Tens to hundreds of tagged objects canbe close together in front of an an-tenna, although objects in use tend tobe further from other objects. Thereader should be able to spot the fewwisps moving at any given time regard-less of the number of tags in the space.

• Most objects tend to move for a few(up to 10) seconds when in use, withsubstantial change in orientation dur-ing the process. Readers should beable to detect orientation changesrobustly within this window.

�-wisp designThe �-wisp uses ID modulation to

communicate its sensor data. Rather thanthe obvious approach of encoding sensordata in special bits reserved in the proto-col, we encode the sensor information incorrelations in the pattern of ID reads.More specifically, we associate each ob-ject with two IDs instead of just one. Wemount two off-the-shelf RFID integratedcircuits on each object. The sensing sub-system controls which of the two IDs isreturned by short-circuiting one or theother of the ID chips. If the first ID isdetected, it indicates that the tagged objectis present and its sensor is in state one; ifthe second ID is detected, we know thatthe object is present and has sensor statetwo. This is equivalent to allocating onebit of ID data to sensing because it causesthe ID space’s size to shrink by one bit.Future wisps will use more sophisticatedform of ID modulation in which multiple

40 PERVASIVEcomputing www.computer.org/pervasive

ENERGY HARVESTING & CONSERVATION

The most common autonomous action studied

so far is prompting users. Our particular interest

is in activity-based reminders.

bits of sensor data are encoded in patternsof ID change over time.

ID modulation lets us add new capa-bilities, such as sensing, while maintain-ing compatibility with existing protocols,standards, and reader hardware. Build-ing the new capabilities as an additionalapplication layer on top of the unmodi-fied preexisting lower layers lets ap-plication-layer capabilities grow anddevelop independently of the lower lay-ers. The investment in these lower layerscan be preserved, while still allowinginnovation at the application layer. Bycontrast, introducing a new protocolwith bit positions reserved for sensor datawould require writing off previous invest-ments in reader hardware as well as newinvestments in replacement reader hard-ware and standards development.

One-bit accelerometerWe’ve created an experimental one-bit

accelerometer that can be read by IDmodulation. It requires no additionalpower supply and thus is a battery-free,wirelessly readable, uniquely identifiedone-bit inertial sensor. It’s compatiblewith the increasingly important 915MHz Electronic Product Code (EPC)standard for RFID. Because the sensorinformation is represented at the appli-cation layer, the system is backwardcompatible with the existing protocol.

We implemented the experimentalsensor using mercury switches. A singlemercury switch can serve as a one-bit,one-axis accelerometer. Suppose theswitch is initially far from any gravita-tional field. The switch has a particularorientation, an axis along which the mer-cury is free to move. When the switchhousing accelerates in one direction, themercury appears (in the switch housingframe of reference) to move in the oppo-site direction because of its inertia. When

the mercury is at one end of its container,the switch is closed; when the mercuryis at the other end, the switch is open.So, the switch will be in one state (closed,say) when the acceleration vector’s pro-jection onto the switch orientation vec-tor is positive. The switch will be in theopposite state (open, in this example)when the sign of the acceleration in theswitch-orientation direction is negative.Because of gravity, accelerometers arealso good tilt sensors: when the object’sorientation changes, the three compo-nents of the gravitational force vectorchange in the object’s frame of reference.Thus, even a one-bit accelerometer func-tions effectively as a tilt sensor.

We can use the switch’s state to mod-ulate the information returned by thetag. If the switch is connected in parallelwith the RFID IC, when the switch isopen, the tag will return its ID normally.When the switch is closed, the tag willbe shorted out and won’t return an ID.If the switch is connected in seriesbetween the RFID chip and the antenna,the ID reads normally when the switch isclosed. Both of these configurations cor-respond to on-off keying (OOK) modu-

lation. (Here, “modulation” refers notto the low-level physical-layer commu-nication protocol but to modulation ofthe entire tag ID presence.)

For many applications, OOK modu-lation is a disadvantage, because distin-guishing “object present, accelerationstate 0” from “object absent” isn’t pos-sible. But using two antiparallel switchesand two RFID ICs, we have imple-mented a binary-code-shift keyingscheme that doesn’t have this problem.Under positive acceleration, the first IDis returned; under negative acceleration,the second ID is returned. If neither ID isreturned, the object is absent.

Binary-code-shift keyed ID modulationcan be implemented in the serial-switchor the parallel-switch configuration. Theserial-switch configuration (which weused for the �-wisp we describe in the“Performance results” section) consistsof two RFID IC chips attached to oneantenna with a mercury switch in serieswith each chip, as Figure 1a shows. Theparallel-switch configuration uses twoIDs and two antennas, with the switchesshort-circuiting one of the IDs. Bymounting the mercury switches in a geo-

JANUARY–MARCH 2005 PERVASIVEcomputing 41

Figure 1. �-wisp serial-switch configuration: (a) schematic, (b) photograph, (c) deployment on a coffee cup.

HgHg HgHg

RFID2

RFID1

Antenna

HgHg

Tag returns ID 1

(b) (c)(a)

metrically antiparallel configuration, weensure that when one switch is open, theother is closed. This way, only one RFIDIC is enabled at any given time. WhichID is returned depends on the object’sacceleration: parallel to switch one cor-responds to ID 1; parallel to switch twocorresponds to ID 2.

VariantsWe have built additional �-wisp vari-

ants. One lets us use just one switch buttwo tags and two chips. We apply anunmodified tag to the object as a pres-ence detector, and use the second tag(with the switch) to encode the sensordata via OOK. Other extensions are pos-sible. We can wire more than two IC-switch pairs in parallel to sense addi-tional states or use more than one switchwith each IC. Wiring the switches in par-allel with each other (and as a group inseries with the chip) causes the ID toreturn when any switch is triggered.Wiring the switches in series with eachother causes the ID to return only whenall the switches are triggered.

A fundamental trade-off of RFID-stylecommunication is between the numberof tags sharing the channel to the reader

and the rate at which each tag can com-municate with the reader. In general, themore tags in the reader’s view field, thelower the communication rate betweeneach tag and the reader. When usingRFID technology to implement densesensor networks (as with wisps), youneed to manage this trade-off in a waythat meets the sensor deployment’s over-all performance requirements. ExistingRFID technology provides built-in meth-ods to mitigate this problem. For instance,RFID readers can put particular tags tosleep so that they don’t use the channelfor an interval of time. In a sensor net-work context, tags can remain quietunless the sensor has significant data toreport. For example, we’ve implementeda variant of the �-wisp that shorts theRFID antenna when the object is in its restposition, so only objects in use—that is,those being moved—consume readerchannel resources.

Performance resultsWe fabricated a serial switch �-wisp

using Alien Technologies’ ALL-9250EPC-compliant UHF RFID tags andSignal Systems International’s Series5008pc mercury switches. We oriented

the switches to be approximately anti-parallel to one another. Figure 1b showsa similar �-wisp.

The RFID tags are 1 cm by 10 cm, andthe switches are roughly 5 mm in lengthand diameter. As Figure 1c shows, an �-wisp fits comfortably on a coffee cup.Furthermore, the tags cost roughly $.40each, and the switches cost $.50. A one-switch, one-tag �-wisp therefore costs$.90; double for the two-switch, two-tagversion used in our tests. Although forper-item RFID tagging to be viable insupply chain applications, tag cost needsto drop much lower, a $1 cost for tag-ging select items is reasonable for thehealthcare and BKM applications we’reconsidering.

Figure 2 shows RFID read rates as afunction of distance. The reader polledthe tag under computer control at ap-proximately 6 Hz (6 polls per second)for about 20 seconds. The read rate isthe ratio of tag responses to readattempts. The top curve is the read ratefor an unmodified RFID tag; the lowercurves are rates for �-wisps in two dif-ferent orientations. We generated thegreen curve (�-wisp, down) by placingthe cup upside-down, which activatesRFID IC 2—the extra IC we added tothe antenna. The red curve representsresponses from the IC that the manu-facturer originally included with theantenna.

Despite the modifications, the tagcontinues to be readable (at a degradedrate) to three meters. At two meters, themodified tag is readable at nearly its fullrate. We conducted this test under some-what ideal circumstances (a metal-freeenvironment and stationary tag), so the

42 PERVASIVEcomputing www.computer.org/pervasive

ENERGY HARVESTING & CONSERVATION

0 0.5 1.0 1.5 2.0 2.5 3.0 3.50

0.2

0.4

0.6

0.8

1.0

Distance (m)

Read

rate

�-wisp, up�-wisp, down

Unmodified tag

Figure 2. The dependence of read rate onread distance. Data is for an �-wisp serial-switch configuration. Read rate is thefraction of reads in which the readerdetected a single stationery tag with noocclusion in typical indoor conditions. �-wisps are detectable beyond 3.0 metersfrom the reader. Read rates are lowerthan for unmodified RFID tags.

particular range values should be viewedas upper bounds for what’s achievablewith this �-wisp implementation. How-ever, we didn’t optimize the implemen-tation itself for range and made no spe-cial effort to minimize the impact ofattaching the sensor to the tag. Further-more, unmodified RFID tags have exhib-ited these ranges even in nonideal con-ditions (such as assembly lines andwarehouses). We expect a more carefulwisp implementation to have nearly thesame communication range as the com-ponent RFID tag. Wisps should there-fore be able to leverage the trend in rangeimprovement of commercial UHF RFIDtags. Recently released tags, for instance,have a nominal range of 6 to 10 meters.At these ranges, we could cover fourrooms (with one antenna each) using asingle $1,500 four-antenna reader.

Figure 3 shows the potential impact ofa worst-case dense deployment on tagdetection rate. The curve’s x-axis is thenumber of unmodified RFID tags evenlyspaced along a 0.6-meter-long paper enve-lope. The antenna reading the tags is 0.9meter away from them. The y-axis is thefraction of placed tags that the readerdetects in a 30-second window. In theabsence of special measures, the fractionof tags detected clearly falls off steeplywith tag density. We emphasize that thisfigure illustrates a worst-case scenario thatis nevertheless quite possible in a wispdeployment. We expect improvementsboth from using different reader configu-rations and from more common lower-density deployments. However, in appli-cations in which very few sensors aresimultaneously active, optimally disablinginactive sensors can be key in enabling ahigh-density deployment.

Figure 4 shows the �-wisp’s operation.We mounted the wisp and a commercialaccelerometer on a coffee cup, which wetilted periodically by motor at 0.18 Hz.The RFID reader provides wireless com-munication and power for the �-wisp; awired USB connection provides commu-nication and power for the STMicro-electronics LIS3L02D accelerometer. Theblue trace shows the acceleration in the x-direction measured by the accelerom-eter incorporated into the Intel Re-search/University of Washington iMotesensor board. The superimposed redtrace is the �-wisp’s output. The red traceis clearly a one-bit quantization of theblue trace. The quantization threshold is nonzero because of the angle at whichwe mounted the �-wisp on the coffee cup.The y-axis label for the red trace is arbi-trary: the +0.8 G and –0.8 G levels shownactually correspond to ID 1 and ID 2.

The thickening of the red trace visibleat some transitions is due to the sensorswitch’s bouncing—that is, as it transi-tions, the switch state can oscillate backand forth before settling to its final newvalue. The host application could “de-bounce” the �-wisp output using a sim-ple software-filtering operation of the

kind commonly used to debounce com-puter keyboard input.

Low-to-high transitions started at 0.14G � 0.08 G (8.2 degrees � 4.4 degrees)and ended at 0.28 G � 0.09 G (16.5degrees � 5.2 degrees). (In other words,we’re reporting the mean and standarddeviation of the distribution of accelera-tion values measured before and after alow-to-high transition.) High-to-lowtransitions started at 0.23 G � 0.06 G(13.0 degrees � 3.5 degrees) and endedat 0.10 G � 0.07 G (5.7 degrees � 4.3degrees). We can’t observe the actualpoint at which the one-bit sensorchanges; we can only see the accelerationvalues immediately before and immedi-ately after the transition. So, the true low-to-high threshold is probably around12.3 degrees (averaging the values beforeand after the transition), and the truehigh-to-low threshold is probably around9.3 degrees. The three-degree differenceis hysteresis and is likely due to imperfectalignment of the (nominally) antiparal-lel mercury switches. The spread in tran-sition values, approximately 0.1 G or 5degrees, is the sensor’s effective angular-resolution limit.

The measurement time resolution—

JANUARY–MARCH 2005 PERVASIVEcomputing 43

2 4 6 8 10 12 140

0.2

0.4

0.6

0.8

1.0

Number of tags in reader field of view

Sing

le-ta

g re

ad ra

te

Figure 3. The effects of a worst-casedense tag deployment. The fraction ofdeployed tags that the reader can read is inversely proportional to the number of tags deployed. The tags are spaceduniformly in a 0.6-meter-wide paperenvelope, read over 30-second windows,at a 0.9-meter range.

that is, the rate at which the PC polledthe reader—is approximately 0.17 sec-onds. The RFID interrogation processappears to be the factor limiting the timeresolution. Some of the apparent accel-eration/angle noise is presumably due tothe time quantization. Angular resolu-tion could likely be improved somewhatby filtering out switch bounces, thoughat some cost in time resolution.

Anatural question is whether wecan generalize the working �-wisp sensor to enable moresophisticated sensing capabil-

ities. The answer is that we can general-ize it and are working to do so.

The mercury switches in the �-wispsupport two logically distinct functions:sensing (the one-bit accelerometer) andcommunication (ID modulation). Inour next generation of wisps, thesefunctions will be cleanly split. Power-harvesting circuitry will drive an ultra-low-power microcontroller with ananalog-to-digital converter that we canconnect to any compatible sensor, sub-ject to power constraints. The micro-controller will perform ID modulationusing an electronic switch, which,unlike the mercury switch, won’t play arole in sensing. By actively modulatingthe pattern of IDs returned over time,the microcontroller will be able to sendarbitrary multibit sensor data, at a lim-

ited rate, through a conventional, stan-dards-compliant RFID reader channel.

Just as in the single-bit sensor datacase, multibit data communicated by IDmodulation requires only two IDs. Thepattern, spread over time, of switchingbetween these two ID values will repre-sent the multibit sensor information. So,moving from single- to multibit sensordata won’t require sacrificing additionalbits of ID space. We will give up just onebit of ID space and gain the ability tosend multiple bits of sensor data throughthe reader channel, at a constrained rate.The generalized platform we envisagecould support multibit measurement ofparameters such as temperature, light,strain, and acceleration.

44 PERVASIVEcomputing www.computer.org/pervasive

ENERGY HARVESTING & CONSERVATION

0 10 20 30 40 50 60

1.0

0.8

0.6

0.4

0.2

0

0.2

0.4

0.6

0.8

1.0

Time (sec.)

�-wispSTMicro LIS3L02D

Acce

lera

tion

(G)

Figure 4. An �-wisp parasitically powered one-bit accelerometer in operation. Comparison of data returned by the ��wisp (the redtrace) with continuous acceleration data provided by an STMicroelectronics LIS3L02D accelerometer (the blue trace).

ACKNOWLEDGMENTSThanks to Jonathan Lester for help

with his iMote Sensor Board.

REFERENCES1. D.E. Culler and H. Mulder, “Smart Sensors

to Network the World,” Scientific Ameri-can, June 2004, pp. 85–91.

2. K. Finkenzeller, RFID Handbook, 2nd ed.,John Wiley & Sons, 2003.

3. E. Dishman, “Inventing Wellness Systems

for Aging in Place,” Computer, vol. 37, no.5, 2004, pp. 34–41.

4. M. Philipose et al., “Inferring ADLs fromInteractions with Objects,” IEEE PervasiveComputing, vol. 3, no. 4, 2004, pp. 50–57.

5. S. Consolvo et al., “Technology for CareNetworks of Elders,” IEEE Pervasive Com-puting, vol. 3, no. 2, 2004, pp. 22–29.

6. K. Fishkin et al., “Ubiquitous ComputingSupport for Skills Assessment in MedicalSchool,” Proc. Ubihealth 2004: The 3rd Int’lWorkshop Ubiquitous Computing for Per-vasive Healthcare Applications, 2004, www.pervasivehealthcare.com/ubicomp2004/papers/final_papers/fishkin.pdf.

JANUARY–MARCH 2005 PERVASIVEcomputing 45

the AUTHORS

Matthai Philipose is a member of the research staff at Intel Research Seattle. Hisresearch interests include statistical reasoning and programming languages. He re-ceived his MS in computer science from the University of Washington, Seattle. Con-tact him at Intel Research Seattle, 1100 NE 65th St., 6th floor, Seattle, WA 98105;matthai.philipose@intel.com.

Joshua R. Smith is a member of the research staff at Intel Research Seattle. Hisresearch focuses on the detection of weak signals for sensing or communicationspurposes. He coinvented electric-field imaging, invented the FiberFingerprint tech-nology, and has published in the field of digital watermarking. He received his PhDfrom the MIT Media Lab's Physics and Media group. He is a member of the IEEE andthe American Association for the Advancement of Science. Contact him at Intel Re-search Seattle, 1100 NE 45th St., Seattle, WA 98105; joshua.r.smith@intel.com.

Bing Jiang is a PhD candidate at the University of Washington. His research interestsinclude radio-frequency-identification technology, sensors, robotics, and power. Hereceived his MS in electrical engineering from the University of Washington, Seattle.Contact him at Dept. of Electrical Engineering, Box 352500, Univ. of Washington,Seattle, WA 98195; bjiang@ee.washington.edu.

Alexander Mamishev is an assistant professor of electrical engineering and director of the Sensors,Energy, and Automation Laboratory (SEAL) at the University of Washington. His research interests includesensor design and integration, dielectrometry, electric power applications, bioengineering, and robot-ics. He received his PhD in electrical engineering from the Massachusetts Institute of Technology. He's amember of the IEEE. Contact him at Dept. of Electrical Engineering, Box 352500, Univ. of Washington,Seattle, WA 98195; mamishev@ee.washington.edu.

Sumit Roy is a professor of electrical engineering at the University of Washington. His research interestsinclude analysis and design of communication systems and networks, with a topical emphasis on next-generation mobile and wireless networks. He received his PhD in electrical engineering from the Univer-sity of California, Santa Barbara. He is a member of the IEEE. Contact him at Dept. of Electrical Engineer-ing, Box 352500, Univ. of Washington, Seattle, WA 98195; roy@ee.washington.edu.

Kishore Sundara-Rajan is a PhD candidate at the University of Washington and anintern at Intel Research Seattle. His research interests include sensor design and inte-gration, on-chip power scavenging, RFID, and dielectric spectroscopy. He receivedhis MS in electrical engineering from the University of Washington, Seattle. He's astudent member of the IEEE. Contact him at Dept. of Electrical Engineering, Box352500, Univ. of Washington, Seattle, WA 98195; kishore@ee.washington.edu.

Tomorrow'sPCs, handhelds,and Internet willuse technologythat exploits current researchin artificial intelligence.Breakthroughs in areas suchas intelligent agents, theSemantic Web, data mining,and natural languageprocessing will revolutionizeyour work and leisureactivities. Read about thisresearch as it happens in IEEE Intelligent Systems.

ww

w.c

om

pu

ter.

org

/in

tell

igen

t/su

bsc

rib

e.h

tm

SEE THEFUTURE OFCOMPUTING

NOWin IEEEIntelligent Systems