8
Lighthouse Localization of Wireless Sensor Networks for Latency-Bounded, High-Reliability Industrial Automation Tasks Felipe M. R. Campos, Craig B. Schindler, Brian G. Kilberg, Kristofer S. J. Pister Berkeley Sensor & Actuator Center Department of Electrical Engineering and Computer Sciences University of California, Berkeley, USA Email: {fmrcampos, craig.schindler, bkilberg, ksjp}@berkeley.edu Abstract—We present the results of a latency-bounded, high- reliability conveyor belt control system for a cart containing a self-localizing wireless sensor node. The node is equipped with an ARM Cortex-M3 microprocessor, 802.15.4 transceiver, 9- axis inertial measurement unit (IMU), and an infrared-sensitive photodiode which allows the wireless node to localize itself using a high-precision localization system for small, resource- constrained, low-cost wireless sensor nodes known as “light- house” localization. The cart moves across the conveyor belt, and upon reaching a specified position sends a wireless signal to a set of receiving nodes attached to the conveyor belt’s motor to reverse direction. Using an extended Kalman filter (EKF) running on-board the cart’s wireless sensor node to estimate the position and velocity of the cart, we are able to achieve 3ms response latency, equivalent to the response latency of industrial photoelectric sensors used in a related implementation. We also show the lighthouse system used in this implementation has no outlier measurements outside the ±1mm error range when stationed 3 meters away from the conveyor belt. This, in addition to use of the EKF, enables high-reliability control with strong occlusion tolerance. We show the wireless sensor node is able to continue estimating its position along the conveyor belt when occluded from the lighthouse base station with a median standard deviation reported by the EKF of 0.875mm after 10 cm of occlusion compared to a median 0.109mm standard deviation of the position estimate when not occluded. Index Terms—Wireless Sensor Networks, Localization, Factory Automation, Industry 4.0, Lighthouse Localization, Extended Kalman Filter, Latency, Reliability, Internet of Things I. I NTRODUCTION Wireless sensor networks (WSNs) provide valuable infor- mation for industrial automation applications that can improve safety, productivity, and efficiency for factory operations. The rise of the Industrial Internet of Things (IIoT) has been and will continue to be driven by the availability of WSNs, embedded systems, and inexpensive sensors [1] [2] [3]. WSNs reduce the need for costly wiring in industrial sensor systems, which is especially important in mobile systems, where mobile wiring harnesses have limited lifetime and increased cost [4], and information from wireless sensor networks will be integral in enabling the implementation of Industry 4.0 [5]. 978-1-7281-5297-4/20/$31.00 ©2020 IEEE Tx Rx Rx Motor Controller Left Right Left Right EKF Position Estimates Transmission from Tx Mote Cart Left Unsafe Zone Right Unsafe Zone Lighthouse Base Station Structured IR Light Fig. 1. Schematic of the conveyor belt wireless industrial control system. The transmitting (TX) wireless sensor node tracks its position with an extended Kalman filter (EKF) as it moves along the cart using acceleration data from its onboard IMU and azimuth measurements calculated from the temporally- structured IR light detected by a photodiode circuit on the wireless sensor mote. Once the cart’s position estimate reaches either “unsafe zone” boundary on the left or right of the conveyor belt, a wireless control signal is sent to the receiving (RX) wireless sensor nodes on separate channels, which tell the motor controller to change the direction of the conveyor belt. Wireless sensor and actuator networks could also enable wireless control in factory automation, taking advantage of the benefits of wireless communication for control. Low- latency, high-reliability networks are of paramount importance for successful automation of wireless factory robots [6]. A. Localization in Industrial Settings For many mobile systems, location information and tracking greatly improves the quality and utility of the data and is vital for them to perform intelligent tasks in industrial environ- ments. This is especially true for wireless industrial control, where position feedback can be critical for the use of the system. RFID has shown promise in tracking assets throughout production, but lacks the reliability or accuracy to be suitable for control applications [7]. Lighthouse positional tracking, a

Lighthouse Localization of Wireless Sensor Networks for

  • Upload
    others

  • View
    4

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Lighthouse Localization of Wireless Sensor Networks for

Lighthouse Localization of Wireless SensorNetworks for Latency-Bounded, High-Reliability

Industrial Automation TasksFelipe M. R. Campos, Craig B. Schindler, Brian G. Kilberg, Kristofer S. J. Pister

Berkeley Sensor & Actuator CenterDepartment of Electrical Engineering and Computer Sciences

University of California, Berkeley, USAEmail: {fmrcampos, craig.schindler, bkilberg, ksjp}@berkeley.edu

Abstract—We present the results of a latency-bounded, high-

reliability conveyor belt control system for a cart containing

a self-localizing wireless sensor node. The node is equipped

with an ARM Cortex-M3 microprocessor, 802.15.4 transceiver, 9-

axis inertial measurement unit (IMU), and an infrared-sensitive

photodiode which allows the wireless node to localize itself

using a high-precision localization system for small, resource-

constrained, low-cost wireless sensor nodes known as “light-

house” localization. The cart moves across the conveyor belt,

and upon reaching a specified position sends a wireless signal to

a set of receiving nodes attached to the conveyor belt’s motor

to reverse direction. Using an extended Kalman filter (EKF)

running on-board the cart’s wireless sensor node to estimate

the position and velocity of the cart, we are able to achieve 3ms

response latency, equivalent to the response latency of industrial

photoelectric sensors used in a related implementation. We also

show the lighthouse system used in this implementation has

no outlier measurements outside the ±1mm error range when

stationed 3 meters away from the conveyor belt. This, in addition

to use of the EKF, enables high-reliability control with strong

occlusion tolerance. We show the wireless sensor node is able

to continue estimating its position along the conveyor belt when

occluded from the lighthouse base station with a median standard

deviation reported by the EKF of 0.875mm after 10 cm of

occlusion compared to a median 0.109mm standard deviation

of the position estimate when not occluded.

Index Terms—Wireless Sensor Networks, Localization, Factory

Automation, Industry 4.0, Lighthouse Localization, Extended

Kalman Filter, Latency, Reliability, Internet of Things

I. INTRODUCTION

Wireless sensor networks (WSNs) provide valuable infor-mation for industrial automation applications that can improvesafety, productivity, and efficiency for factory operations. Therise of the Industrial Internet of Things (IIoT) has beenand will continue to be driven by the availability of WSNs,embedded systems, and inexpensive sensors [1] [2] [3]. WSNsreduce the need for costly wiring in industrial sensor systems,which is especially important in mobile systems, where mobilewiring harnesses have limited lifetime and increased cost [4],and information from wireless sensor networks will be integralin enabling the implementation of Industry 4.0 [5].

978-1-7281-5297-4/20/$31.00 ©2020 IEEE

Tx

Rx

Rx

Motor Controller

Left

Right

Left

Right

EKF Position Estimates

Transmissionfrom Tx Mote

Cart

Left Unsafe Zone Right Unsafe ZoneLighthouseBase Station

StructuredIR Light

Fig. 1. Schematic of the conveyor belt wireless industrial control system. Thetransmitting (TX) wireless sensor node tracks its position with an extendedKalman filter (EKF) as it moves along the cart using acceleration data fromits onboard IMU and azimuth measurements calculated from the temporally-structured IR light detected by a photodiode circuit on the wireless sensormote. Once the cart’s position estimate reaches either “unsafe zone” boundaryon the left or right of the conveyor belt, a wireless control signal is sent tothe receiving (RX) wireless sensor nodes on separate channels, which tell themotor controller to change the direction of the conveyor belt.

Wireless sensor and actuator networks could also enablewireless control in factory automation, taking advantage ofthe benefits of wireless communication for control. Low-latency, high-reliability networks are of paramount importancefor successful automation of wireless factory robots [6].

A. Localization in Industrial Settings

For many mobile systems, location information and trackinggreatly improves the quality and utility of the data and is vitalfor them to perform intelligent tasks in industrial environ-ments. This is especially true for wireless industrial control,where position feedback can be critical for the use of thesystem. RFID has shown promise in tracking assets throughoutproduction, but lacks the reliability or accuracy to be suitablefor control applications [7]. Lighthouse positional tracking, a

Page 2: Lighthouse Localization of Wireless Sensor Networks for

localization system that uses a temporally-structured infraredillumination scheme generated by a set of base stations toconvey position information, is another technology that hasshown high precision in industrial applications [8]. For in-stance, Nikon’s large-scale metrology system, iGPS, advertises200 micrometer precision [9]. HTC Vive, a consumer virtualreality (VR) headset and corresponding lighthouse motioncapture system, has also been gaining significant attention asa cost-effective, high-precision pose tracking system that usesa method similar to Angle-of-Arrival (AOA) and advertisesmillimeter precision [10] [11].

Additionally, reliability and robustness are important forreal-time industrial control systems. Various environmentalfactors in an industrial setting, such as object occlusions inoptical localization systems, can adversely affect the accuracyof positioning measurements. One way to improve robustnessand reliability is by tracking state estimations based on inertialmeasurements and previous position measurements. This canreduce the effect of erroneous measurements from one ofthe sensors by intelligently combining them with independentmeasurements from other sensors. One of the most commonways of doing this is with the addition of a built-in inertialmeasurement unit (IMU) and sensor fusion algorithms like theKalman filter [12]. Most position tracking technologies useeither small, passive markers, like RFID or motion capture, oruse expensive and large tracking beacons, such as iGPS. Theprohibitive cost and size of active wireless tracking beaconsmakes them unsuitable for ubiquitous applications where theremay be hundreds of sensors.

B. Implementation and Related WorkOur work focuses on using a miniature, wireless inertial

sensor mote to implement a low-latency, high-precision, high-reliability state estimation and wireless control system fortracking the position of objects in industrial settings. Thereduced size and cost of the measurement system makes itmore amenable to large-scale implementation.

Previously, Schindler et. al. demonstrated performing thissame control task using Azbil HP-model photoelectric sensorsplaced adjacent to the conveyor belt as beam breakers tosense when the cart has reached an “unsafe zone” boundary[13]. Once the beam breakers detect the cart, they trigger aninterrupt on an attached transmitting wireless sensor node,which sends a wireless signal to a motor controller to changethe direction of the conveyor belt. Their implementation (aswith ours) took advantage of a high-reliability, low-latencycommunication protocol from [14] for the IEEE 802.15.4standard with multiple low-cost motes, channel frequencydiversity, and transmission redundancy. This ensures lowtransmission latency and high packet delivery ratio (PDR)to provide the system with a high level of reliability in anindustrial environment.

In our implementation (shown in figure 1) we use an ex-tended Kalman filter (EKF) running on-board a wireless sensornode on the cart for position and velocity estimation usingboth inertial measurements and lighthouse localization. We

(a)

(b)

(c)

Fig. 2. *Figure reprinted from [15]. (a) The HTC Vive lighthouse uses line-generating optics to perform planar sweeps of infrared light across the trackingrange. (b) Once the rotating laser reaches a reference bearing (�⇡

2 for thepurposes of our experiment such that the center of the conveyor belt is 0), anarray of wide-angle LEDs flash for a specified amount of time to sync thetimers of all nodes in the tracking range and communicate (via the duration ofthe pulse) the axis on which the rotating laser is sweeping. (c) Once the planarlaser sweep hits the photodiode mounted on the tracked node, the tracked nodeconverts the time difference between the sweep hit and the original sync pulseinto an azimuth or elevation angle relative to the reference bearing in thelighthouse base station’s reference frame.

show that the EKF provides equivalent response latency to theAzbil sensors when the cart reaches an unsafe zone boundary.We also show that the EKF enables high-reliability controlwith strong tolerance to occlusion from external mountedsensors. Conversely, the Azbil photoelectric sensors used inSchindler et. al.’s system provide no occlusion tolerance orstate estimation as they can only rely on a single measurementfrom the sensor and would not be able to predict when theobject would cross the conveyor belt’s boundary into an unsafezone.

II. PLATFORM & METHOD

A. Lighthouse Localization

For the purposes of this conveyor belt control system, weused HTC Vive’s lighthouse base station, which is used asthe basis for localizing objects in their VR system, and isconsidered a cost-effective, high-precision means of ground

Page 3: Lighthouse Localization of Wireless Sensor Networks for

Fig. 3. The conveyor belt wireless control setup contains (a) a DC regulated power supply, (b) 2 RX MIMSYs attached to (c) an Arduino Uno controlling(d) a DC motor, (e) an HTC Vive lighthouse base station, and (f) a TX MIMSY, powered by a LiPo battery (not pictured), equipped with (g) an MPU-92509-axis IMU and (h) a TS3633-CM1 integrated circuit & infrared-sensitive photodiode for picking up pulse-train envelopes from the lighthouse base station.

truth pose tracking [10] [16]. Vive’s lighthouse, using horizon-tally and vertically rotating planar laser sweeps with omnidi-rectional flashes of infrared light (“sync” pulses) in between,allows for the tracked object to determine its horizontal andvertical angles relative to the base station as shown in figure2. This “inside-out” pose tracking system, efficiently scalablewith the number of localized nodes, additionally provides ob-jects with factory-calibrated constellations of infrared-sensitivediodes on them with the ability to recover their 6 degree-of-freedom (6DOF) pose with millimeter-scale precision atupdate rates of up to 60 Hz [10] [16] [17].

One lighthouse can recover the angular portion of sphericalcoordinates defining the tracked object’s position in the light-house’s reference frame, as shown in figure 2. The horizontal(azimuth) and vertical (elevation) angles (�, ✓ respectively)can be calculated knowing the angular velocity of the planarsweeps, ! = 120⇡ rad

s , and the time difference between thesync pulse and the time the planar laser sweep reaches thetracked object’s photodiode, �t. In our case, this angle-of-arrival (AOA) information is captured using rising and fallingedge interrupts on a wireless sensor node capable of detectinginfrared light. Such that the center of the conveyor belt is at0 rad, we set our reference bearing to �⇡

2 :

� = �th · ! � ⇡

2

✓ = �tv · ! � ⇡

2

(1)

For the purposes of this experiment and for simplicity, weonly require the horizontal position of the cart on the conveyorbelt, and therefore only make use of the horizontal sweep

of the base station module, though the lighthouse systemeasily extends to 3D and 6DOF pose estimation [10] [17].We additionally align the baselines of the lighthouse and theconveyor belt, which at ` meters apart and angle � of the cartrelative to the lighthouse base station’s normal vector, givesthe horizontal position of the cart as:

x = ` · tan(�) (2)

B. Conveyor Belt System

The core of the latency-bounded, high-reliability wirelesslocalization and conveyor belt motor control system is theMicro Inertial Measurement System (MIMSY), an open-source wireless sensor mote containing an ARM Cortex-M3 microprocessor, an InvenSense MPU-9250 9-axis IMU,and 802.15.4 transceiver [13]. Our conveyor belt trackingsystem, shown in figure 3, consists of the following primarycomponents: an infrared-laser sweeping HTC Vive base stationmodule, three MIMSYs, a TS3633-CM1 prototyping module,a DC motor and power supply, and an Arduino Uno.

Using the unofficial documentation on Vive’s basestation and its communication protocols (available atgithub.com/nairol/LighthouseRedox), we wereable to write custom firmware on top of the existing low-latency, high-reliability communication firmware from [14]allowing MIMSY to localize itself on the conveyor belt relativeto the lighthouse. The TS3633-CM1 prototyping module is anintegrated circuit and photodiode that generates an envelopepulse train of the infrared signals received from the lighthouse,which the MIMSY receives as rising and falling edges through

Page 4: Lighthouse Localization of Wireless Sensor Networks for

its GPIO pins. The written firmware uses rising and fallingedge hardware capture registers to record pulse timings preciseto 100 nanoseconds in addition to falling edge interruptsto read from those registers and decode the pulse trains ashorizontal and vertical angle information.

The miniature conveyor belt system shown in figure 3 wasassembled to emulate a factory robot on which to conductlatency and reliability experiments using self-localizing MIM-SYs. The transmitting (TX) MIMSY updates its position andvelocity estimate on the conveyor belt via an extended Kalmanfilter (EKF), taking acceleration data given by MIMSY’sonboard IMU in addition to azimuth data from the lighthousegenerated from the aforementioned hardware capture register& interrupt scheme. As the TX MIMSY approaches a givenedge of the cart, it sends several redundant packets on twodifferent channels, 16 (2.430 GHz) and 20 (2.450 GHz),indicating proximity to the edge of the conveyor belt. The RXMIMSYs receive these packets and send a signal over GPIOpins notifying the motor controller to change the direction ofthe conveyor belt such that the cart doesn’t enter the “unsafezones” shown in figure 1.

As in Schindler et. al.’s implementation, mesh networking orTSCH capabilities of OpenWSN were not used in this system.However, it did build upon the open source OpenWSN boardsupport package (BSP) for the CC2538 system-on-chip (SoC),which provides an abstraction of the underlying hardware [13][18] [19]. Additionally, taking advantage of this lower-levelnetwork control allows for reduced latency when compared tothe full OpenWSN stack implementation, at the expense ofhigher power consumption and limited networking functional-ity [13]. All relevant low-latency, lighthouse localization, andstate estimation firmware written for MIMSY can be found onGithub [20].

C. Extended Kalman Filter

Taking advantage of a diverse set of sensors under aBayesian state estimation scheme directly increases both thereliability and tracking accuracy of our system, simultaneouslyreducing the response latency of control updates. Becauseof the weak nonlinearity introduced by the angle-to-positionmapping (2) in our localization system, we used an extendedKalman filter (EKF) to combine accelerometer measurementsfrom the MPU-9250 IMU with lighthouse azimuth measure-ments.

The dynamics used for state estimation are given in (3). Thesystem states are cart position x(k) and cart velocity v(k).The input is the acceleration measured from the IMU, whichconsists of the true acceleration a and additive noise na withzero mean and variance �

2a. The acceleration is sampled with

period �ta.

x(k + 1)v(k + 1)

�=

1 �ta

0 1

� x(k)v(k)

�+

0

�ta

�(a(k)+na(k)) (3)

The measurement model is given in (4). The measurementz(k) is equal to the sum of the true horizontal angle �(k)

between the lighthouse and the cart, and additive noise n�(k)with zero mean and variance �

2�.

z(k) = h(

x(k)v(k)

�) := �(k) + n�(k) (4)

where:

�(k) = arctan(x(k)

`) (5)

In each step of our recursive EKF, we first predict the nextstate via our dynamics model (6) and corresponding Jacobianmatrix (7):

fk =

x(k)v(k)

�=

1 �ta

0 1

� x(k � 1)v(k � 1)

�+

0

�ta

�a(k�1) (6)

Fk =

1 �ta

0 1

�(7)

We then predict our state covariance (8), with process noisecovariances (9):

Pk = Fk�1Pk�1FTk�1 + Lk�1Qk�1L

Tk�1 (8)

Qk�1 =

0 00 �

2a

�Lk�1 =

0 00 �ta

�(9)

Given a new measurement, zk, the next step is computingthe innovation from said measurement:

zk = zk � h(

x(k)v(k)

�) (10)

We then compute the Kalman Gain (11) from our statecovariance, the Jacobian of our measurement model (12) andour measurement covariance (13):

Kk = PkHTk (HkPkH

Tk +R)�1 (11)

where:

Hk =

2

4`

`2 + x(k)2

0

3

5T

(12)

R = �2� (13)

The next step is to update our state estimate:

xk = xk +Kkzk (14)

Finally, we update our state covariance:

Pk = (I �KkHk)Pk (15)

Page 5: Lighthouse Localization of Wireless Sensor Networks for

Median Overshoot � (SD) Median Latency � (SD)Lighthouse Only 9.9mm 10.9mm 26.7ms 29.5msLighthouse EKF 1.1mm 0.8mm 3.0ms 2.3msAzbil HP Sensors – – 1-3ms* –

Fig. 4. In each plot, the purple dots indicate the time and position at which the TX mote sent a wireless control signal change motor direction. The DCpower supply output voltage was 14V, corresponding to a conveyor belt speed of 36.91cm/s. Left: Overshoot of lighthouse wireless control signal. The medianovershoot was 9.9mm corresponding to an median response latency of 26.7ms. Right: Overshoot of EKF wireless control signal. The average overshoot was1.1mm, corresponding to an average response latency of 3ms. Table: Comparison of overshoot and response latency between the lighthouse-only controlsystem, lighthouse EKF control system, and the Azbil HP model photoelectric sensors used in Schindler et. al.’s implementation of the conveyor belt controlsystem [13]. The lighthouse EKF control system achieves comparable response latency to that of the Azbil HP model photoelectric sensors, and this responselatency could be further decreased by increasing the MPU-9250 accelerometer update rate from the current 1kHz to 4kHz maximum update rate. *Responselatency reported by the Azbil General-Purpose Self-Contained Photoelectric Switches datasheet [21].

III. EXPERIMENTAL RESULTS & DISCUSSION

Schindler et. al. showed that the low-latency board supportpackage (BSP) protocol used for this conveyor belt controlsystem – in taking advantage of channel frequency diversityand transmission redundancy – provides a high level of net-work reliability, both through latency and packet delivery ratio(PDR) [13] [22]. For monitoring network latency, a computerrunning a custom Python test harness and an Analog Discovery2 were used [14] [23]. Over 10,000 packets sent, the maximumlatency was 1.6ms and the minimum latency was 0.8ms, withall 10,000 packets received by the RX motes, demonstratinglow-latency and high-reliability [13].

Experiments on our implementation of the conveyor beltcontrol system were performed with two means of estimat-ing the cart’s position on the conveyor belt, one using thelighthouse alone and one using the EKF described previously.With the lighthouse alone, the TX MIMSY makes use of the60 Hz update rate of individual azimuth measurements fromthe lighthouse base station to infer the position of the cart onthe conveyor belt. With the use of the EKF, the TX MIMSYintelligently combines lighthouse azimuth measurements andacceleration measurements from the IMU in order to estimatethe position and the velocity of the cart at all times whiletaking into account the relative variance of each state estimatethrough time.

Later in this section, we additionally consider the casewhere the conveyor belt may be occluded from the lighthousebase station such that the TX MIMSY can only use the ac-celerometer and previous lighthouse measurements to estimateits state on the belt. In all cases, state-estimation calculationis performed on-board the TX MIMSY in real-time. The TXmote sends the control signal as a 10 byte packet: the first fourbytes containing a unique passphrase that identifies the TXmote, the fifth byte communicating which side of the conveyor

belt the cart has reached, and the remaining bytes used for thecurrent position estimate at that time to the RX mote. Velocityestimates, state covariance, and other metadata were sent inplace of the position estimate as necessary.

For the purposes of these experiments, the distance `

between the cart and the lighthouse base station was care-fully calibrated to 3 meters, and the distance between theunsafe zone boundaries on the conveyor belt was 20 cm.The lighthouse azimuth measurement standard deviation, ��,in the EKF was set to 3.5 ⇥ 10�5 radians, and accelerationmeasurement standard deviation, �a, was set to 0.1m/s

2 withan update frequency, 1

�ta, of 1kHz.

A. System Latency

When performing wireless control with the lighthouse alone,for every individual azimuth measurement from the basestation, the TX MIMSY checks that it is within the boundsof the conveyor belt’s boundaries. If measurement indicatesthat the cart has reached or passed either boundary into anunsafe zone, a wireless control signal is transmitted and theRX motes change the direction of the conveyor belt.

HTC Vive’s lighthouse’s angle measurement update rate of60 Hz means that for a cart travelling at 1 m/s, the cart travels1.67 centimeters between lighthouse measurement updates,such that the TX MIMSY may not update its position thatthe cart has entered an unsafe zone until it has travelled1.67 centimeters inside it. Using the extended Kalman filter(EKF) allows us to take advantage of all previous positionmeasurements from the lighthouse, using the accelerometer topredict the state in between lighthouse measurement updates,for a continuously updating position and velocity estimate,mitigating any issues with the system’s response latencybrought on by the lighthouse’s 60 Hz update rate.

Page 6: Lighthouse Localization of Wireless Sensor Networks for

Fig. 5. Lighthouse azimuth measurement distribution over 1 million sampleswith the base station 3 meters away from the conveyor belt. The mean of thedistribution was centered to zero and the standard deviation was 3.5⇥ 10�5

rad, equivalent to ⇡ 100µm or .1mm at 3 meters away. The red error bars oneither side of the sampled distribution indicate ±1 mm error of the lighthousemeasurement. Over 1 million sampled azimuth points, 0% were outside the±1 mm error range (marked with vertical red dotted lines).

Figure 4 shows how the use of the EKF improves the re-sponse latency of the wireless control system by continuouslypredicting the position and velocity states of the cart betweenupdates from the lighthouse sensor. The EKF implementationupdating at a 1kHz accelerometer update rate enables a medianresponse latency of 3.0ms, a significant improvement onthe 26.7ms median response latency when using lighthousemeasurements alone, and an equivalent response latency tothe 1-3ms latency of Azbil’s HP model photoelectric sensorsused in Schindler et. al.’s implementation [13] [21]. The EKF’sresponse latency could be further improved by increasing theMPU-9250 accelerometer update rate from the current 1kHzto 4kHz maximum update rate reported by the datasheet [24].

B. System Reliability & Occlusion ToleranceAs an initial metric of the lighthouse measurement system’s

reliability, we quantify the precision of the lighthouse hori-zontal angle measurements on a MIMSY using its DMA &interrupt scheme for capturing infrared pulses and sweeps. Infigure 5, we display the distribution of 1 million measurementstaken, which reports a standard deviation of 35 microradians,equivalent to 100 microns or .1mm with the base station at adistance of 3 meters away. Additionally, we see zero outliersoutside of the ±1 mm error range at 3 meters away over 1million measurements, further reinforcing the precision andreliability of this lighthouse localization system.

As with the photoelectric sensors used in Schindler et. al.,use of the lighthouse alone provides no tolerance to occlusionsince lighthouse measurements are not received when the light-house base station emitter is occluded from the photodiodeon the TX MIMSY. In the case that occlusion occurs at aboundary of an unsafe zone, the TX MIMSY is unaware that it

Fig. 6. Results from a set of experiments where we occluded the entire lefthalf of the conveyor belt (including the unsafe zone) from the lighthousebase station in an attempt to analyze our system’s tolerance to occlusion. No

Occlusion (Left): EKF position estimate variance distribution at the unsafezone boundary when no part of the belt is occluded from the lighthousebase station’s infrared sweeps. The position estimate at the conveyor beltboundaries has a median variance of 1.193 ⇥ 10�8m2, corresponding toa median standard deviation of 0.109mm. 10cm Occlusion (Right): EKFposition estimate variance distribution at the unsafe zone boundary in thecase where half of the belt is occluded from the base station. The positionestimate has a median variance of 7.659⇥10�7m2 at the occluded boundary,corresponding to a median standard deviation of 0.875mm after 10cm ofocclusion on the belt, indicating tolerance to occlusion and, by extension,significant additional reliability in our system.

has reached a boundary and will enter the unsafe zone withoutsending a wireless control signal. Accelerometers alone areknown for producing notably poor translation estimates fordead-reckoning over long periods of time due to scale factorsand bias drift [25] [26]. However, in our EKF implementation,information from all previous lighthouse measurements arepropagated into future states even once occluded such thatthe accelerometer can be successfully used to perform precisepositional tracking while occluded from the lighthouse basestation.

We performed a set of experiments where we occluded theentire left half of the conveyor belt (including the unsafe zoneboundary) from the lighthouse base station in an attempt toanalyze our system’s tolerance to occlusion. Figure 6 showsthe results of this experiment, displaying the EKF positionestimate variance distributions over several cycles back andforth across the conveyor belt. When no part of the conveyorbelt is occluded from the lighthouse base station’s infraredsweeps, the position estimate at the unsafe zone boundarieshas a median standard deviation of 0.109mm. In the casewhere half of the conveyor belt is occluded from the basestation, the EKF reports a median position estimate standarddeviation at the occluded boundary of 0.875mm after 10cmof occlusion on the belt, indicating tolerance to occlusion and,by extension, significant additional reliability in our system.

Page 7: Lighthouse Localization of Wireless Sensor Networks for

IV. CONCLUSION

The goal of this work was to quantify the latency and relia-bility of a conveyor belt control system for a cart containing atransmitting Micro Inertial Measurement System (MIMSY).The MIMSY on the cart receives pulse-timing informationfrom an infrared-sensitive photodiode, allowing the MIMSY toperform wireless position control on the belt while localizingitself relative to an HTC Vive lighthouse base station placed3 meters perpendicular from the conveyor belt.

Using an extended Kalman filter (EKF) running on-boardthe TX MIMSY, we were able to achieve 3ms response latencyto the cart reaching an unsafe zone boundary on the belt,equivalent to the response latency of Azbil HP model industrialphotoelectric sensors used in Schindler et. al.’s implementationof the conveyor belt control system. We also showed that at3 meters away, the lighthouse localization system affordedby HTC Vive’s lighthouse base station obtains sub-millimeterprecision localization with minimal outliers. This in addition touse of the EKF, which takes into account the relative varianceof lighthouse and accelerometer measurements through timeto estimate the position and velocity of the cart, enabledhigh-reliability control with strong occlusion tolerance. This isexemplified by the TX MIMSY’s ability to continue estimatingits position on the conveyor belt with millimeter precisioneven after 10cm of occlusion from the lighthouse base station.Conversely, with the HP model photoelectric sensors used inSchindler et. al.’s control system, we would only be able torely on a single measurement from the sensor and would notbe able to predict when the object would cross the conveyorbelt’s boundary into the unsafe zone.

V. FUTURE WORK

Though for this work we used the Micro Inertial Measure-ment System (MIMSY) as the wireless sensor node at the coreof the system, this “inertial lighthouse” wireless control systemis directly compatible with a monolothic wireless system-on-chip (SoC) that requires no external components, known asthe Single Chip Micro Mote (SCµM) [27]. SCµM contains anARM Cortex-M0 microprocessor and a 2.4GHz radio compat-ible with Bluetooth Low Energy (BLE) and the IEEE 802.15.4standard, the latter allowing us to implement the same low-latency board support package (BSP) implemented for the TXMIMSY to communicate to the RX motes. SCµM additionallycontains an optical receiver whose primary function is foroptical bootloading. Additionally, this same optical receiveris able to detect the infrared pulses emitted by the HTC Vivelighthouse base stations in order to localize itself [15], and hasbeen shown to achieve centimeter precision localization in atwo lighthouse system [28]. Future work could additionallyinvestigate compatibility with Nikon’s large-scale metrologyiGPS system [9]. This Single Chip Micro Mote will enablesuch cubic millimeter sensor motes to act as low-cost markersfor wireless industrial control systems, in addition to a slew ofother applications in the realm of localization, wireless sensornetworks, and the realization of Industry 4.0.

VI. ACKNOWLEDGEMENTS

This work was supported by the Berkeley Sensor andActuator Center and the UC Berkeley Swarm Lab.

REFERENCES

[1] J. Wan, S. Tang, Z. Shu, D. Li, S. Wang, M. Imran, and A. V.Vasilakos, “Software-Defined Industrial Internet of Things in the Contextof Industry 4.0,” IEEE Sensors Journal, vol. 16, no. 20, pp. 7373–7380,Oct 2016.

[2] S. Wang, J. Wan, D. Zhang, D. Li, and C. Zhang, “Towards SmartFactory for Industry 4.0: A Self-Organized Multi-Agent System withBig Data based Feedback and Coordination,” Computer Networks, vol.101, pp. 158–168, 2016.

[3] S. Wang, J. Wan, D. Li, and C. Zhang, “Implementing Smart Factory ofIndustrie 4.0: An Outlook,” International Journal of Distributed SensorNetworks, vol. 12, 2016.

[4] Imad Jawhar, Nader Mohamed, and Khaled Shuaib, “A framework forpipeline infrastructure monitoring using wireless sensor networks,” in2007 Wireless Telecommunications Symposium, April 2007, pp. 1–7.

[5] Y. Lu, “Industry 4.0: A survey on technologies, applications and openresearch issues,” 2017.

[6] S. A. Ashraf, I. Aktas, E. Eriksson, K. W. Helmersson, and J. Ansari,“Ultra-reliable and low-latency communication for wireless factoryautomation: From LTE to 5G,” in Emerging Technologies and FactoryAutomation (ETFA), 2016 IEEE 21st International Conference on.IEEE, 2016, pp. 1–8.

[7] J. Chongwatpol and R. Sharda, “RFID-enabled track and traceabilityin job-shop scheduling environment,” European Journal of OperationalResearch, vol. 227, no. 3, pp. 453 – 463, 2013. [Online]. Available:http://www.sciencedirect.com/science/article/pii/S0377221713000362

[8] K. Romer, “The lighthouse location system for smart dust,” in Proceed-ings of the 1st international conference on Mobile systems, applicationsand services. ACM, 2003, pp. 15–30.

[9] Nikon, “iSpace: Large Volume Metrology, tracking,and positioning,” Brochure, 2010. [Online]. Available:https://www.nikonmetrology.com/en-gb/product/igps

[10] M. Borges, A. C. Symington, B. Coltin, T. Smith, and R. Ventura,“HTC Vive: Analysis and Accuracy Improvement,” 2018 IEEE/RSJInternational Conference on Intelligent Robots and Systems (IROS), pp.2610–2615, 2018.

[11] A. Paul and T. Sato, “Localization in Wireless Sensor Networks:A Survey on Algorithms, Measurement Techniques, Applications andChallenges,” Journal of Sensor and Actuator Networks, vol. 6, p. 24, 102017.

[12] R. E. Kalman, “A new approach to linear filtering and predictionproblems,” Journal of basic Engineering, vol. 82, no. 1, pp. 35–45,1960.

[13] C. B. Schindler, D. S. Drew, B. G. Kilberg, F. M. R. Campos, S. Yanase,and K. S. J. Pister, “MIMSY: The Micro Inertial Measurement Systemfor the Internet of Things,” in 2019 IEEE 5th World Forum on Internetof Things (WF-IoT) (WF-IoT 2019). Limerick, Ireland: IEEE, Apr.2019.

[14] B. Kilberg, C. B. Schindler, A. Sundararajan, A. Yang, and K. S. Pister,“Experimental Evaluation of Low-Latency Diversity Modes in IEEE802.15. 4 Networks,” in 2018 IEEE 23rd International Conference onEmerging Technologies and Factory Automation (ETFA), vol. 1. IEEE,2018, pp. 211–218.

[15] B. Wheeler, A. Ng, B. Kilberg, F. Maksimovic, and K. S. J. Pister, “Alow-power optical receiver for contact-free programming and 3d local-ization of autonomous microsystems,” in appeared in IEEE UEMCON2019. IEEE, 2019.

[16] D. C. Niehorster, L. Li, and M. Lappe, “The Accuracy and Precisionof Position and Orientation Tracking in the HTC Vive Virtual RealitySystem for Scientific Research,” in i-Perception, 2017.

[17] Y. Yang, D. Weng, D. Li, and H. Xun, “An Improved Method of PoseEstimation for Lighthouse Base Station Extension,” Sensors, vol. 17,no. 10, p. 2411, 2017.

[18] T. Watteyne, X. Vilajosana, B. Kerkez, F. Chraim, K. Weekly, Q. Wang,S. Glaser, and K. Pister, “OpenWSN: a standards-based low-powerwireless development environment,” Transactions on Emerging Telecom-munications Technologies, vol. 23, no. 5, pp. 480–493, 2012.

Page 8: Lighthouse Localization of Wireless Sensor Networks for

[19] CC2538 System-on-Chip Solution for 2.4-GHz IEEE 802.15.4 andZigBee/ZigBee IP Applications, Texas Instruments, 5 2013, version C.

[20] B. G. Kilberg and F. M. R. Campos, “fork of‘OpenWSN Firmware: Stuff That Runs on a Mote’,”https://github.com/skrillberg/low latency fw/tree/ekf, 2019.

[21] General-Purpose Self-Contained Photoelectric Switches, Azbil, 8 2018,7th Edition.

[22] T. Watteyne, A. Mehta, and K. Pister, “Reliability through frequencydiversity: why channel hopping makes sense,” in Proceedings of the 6thACM symposium on Performance evaluation of wireless ad hoc, sensor,and ubiquitous networks. ACM, 2009, pp. 116–123.

[23] A. Yang, A. Sundararajan, C. B. Schindler, and K. S. Pister, “Analysisof low latency TSCH networks for physical event detection,” in WirelessCommunications and Networking Conference Workshops (WCNCW),2018 IEEE. IEEE, 2018, pp. 167–172.

[24] MPU-9250 Product Specification, InvenSense, 6 2016, rev. 1.1.[25] P. Batista, C. Silvestre, P. Oliveira, and B. Cardeira, “Accelerometer

calibration and dynamic bias and gravity estimation: Analysis, design,and experimental evaluation,” IEEE Transactions on Control SystemsTechnology, vol. 19, no. 5, pp. 1128–1137, Sep. 2011.

[26] M. Nießner, A. Dai, and M. Fisher, “Combining Inertial Navigation andICP for Real-time 3D Surface Reconstruction,” Eurographics, pp. 1–4,2014.

[27] F. Maksimovic, B. Wheeler, D. C. Burnett, O. Khan, S. Mesri, I. Suciu,L. Lee, A. Moreno, A. Sundararajan, B. Zhou et al., “A Crystal-Free Single-Chip Micro Mote with Integrated 802.15.4 CompatibleTransceiver, sub-mW BLE Compatible Beacon Transmitter, and CortexM0,” in 2019 Symposium on VLSI Circuits. IEEE, 2019, pp. C88–C89.

[28] B. G. Kilberg, F. M. R. Campos, F. Maksimovic, and K. S. J. Pister,“Accurate 3D Lighthouse Localization of a Low-Power Crystal-FreeSingle Chip Mote,” in to appear in Solid-State Sensors, Actuatorsand Microsystems Workshop (Hilton Head). Transducers ResearchFoundation (TRF), 2020.