Developing Active Sensor Networks with Micro Mobile Robots: Distributed Node Localization

Transcription

1 Developing Active Networks with Micro Mobile Robots: Distributed Node Localization Weihua Sheng and Gira Tewolde Kettering University/Forerly GMI Flint, MI 484 Eail: Abstract Top-down Service Request node A new sensor network architecture called active sensor network (ASN) is proposed in this paper. By integrating ultiple sall, sensor network-friendly obile robots into a traditional sensor network, a closed-loop, dynaic, adaptive sensor network is fored. Such sensor networks have the following erits: adaptivity, self healing, responsiveness and longer lifetie. This paper focuses on the distributed sensor node localization using ultiple obile robots. A potentialbased robot area partition algorith and a localization algorith are developed. Siulation results verify the proposed algoriths. Motivation Introduction Recent advanceents in wireless counication and icro-electro-echanical systes (MEMS) have ade possible the deployent of wireless sensor networks for any real world applications, such as environental onitoring, search and rescue, ilitary surveillance, and intelligent transportation (Akyildiz et al. 2; Mainwaring et al. 2; Siic & Sastry 3), etc. In any situations, people do not have the luxury to carefully distribute the sensors. Instead, the sensors are deployed in large quantities very quickly. For exaple, in environental onitoring and ilitary surveillance, sensors can be dropped fro airplanes. At disaster sites, search and rescue sensor networks are anually deployed by rescue workers in a quick fashion. However, the inherited probles with such traditional wireless sensor networks are: (1) It is very difficult to control the sensor density, coverage and connectivity of the sensor networks. For instance, in environental onitoring applications, certain areas of interest ay require higher sensor density in order to provide ore detailed sensing inforation and such requireents ay arise in a dynaic fashion which can not be predicted beforehand. (2) Once deployed, the sensor network can not adapt to a changing environent. For exaple, the quick change of battlefield situations ay require a surveillance sensor network be redistributed to cover new areas of interest. (3) The overall life tie of the Copyright c 6, Aerican Association for Artificial Intelligence ( All rights reserved. Control Center Robot Inforation Sensing Inforation Botto-up Service Request uservbot Inter-robot Counication Counication Figure 1: An active sensor network: closing the loop through actuation. sensor network is liited by the capacity of the batteries carried by the sensors and the network will not be able to carry out the ission if a critical nuber of sensors deplete their batteries. (4) There lack efficient ethods to quickly deterine the geographic locations of a large nuber of sensors while the location inforation is very iportant to ost of the applications. In this work, contrary to traditional open loop, passive sensor networks, we develop a new sensor network architecture called active sensor network, which eploys ultiple sensor-network-friendly icro service robots (µservbots) to ipleent an actuation echanis and thus closes the loop. Based on a set of core functions, the icro service robots can provide both logistic and network services. Exaples of the logistic services ay include (1) sensing coverage control; (2) sensor power supply and (3) sensor calibration, etc. Exaples of the network services ay include (1) network connectivity, or topology anageent; (2) hierarchical routing and (3) tie synchronization, etc. When these icro service robots are deployed together with a large quantity of sensors, the resulted active sensor network will achieve any desirable erits, such as adaptivity, self healing, responsiveness and longer lifetie. The active sensor network architecture is illustrated in Figure 1, where logistic or network service requests are either initiated by the control center or by the sensors. Here the control center, which can be a solider, a firefighter, a rescue worker, or siply a coputer, is not just an inforation sink (Akyildiz et al. 2) as in any traditional open loop sensor networks. Instead, it can actively generate coands to control and anage the underlying sensor network. 478

2 Related work Ebedding obile robots in sensor networks has received soe attention recently. LarMaca et al. (Laraca et al. 2) proposed a sensor network that uses a robot to carry out the following functions: deploy and calibrate sensors, detect and react to sensor failures. Corke et al. (Corke et al. 4) used a UAV (an autonoous helicopter) to quickly deploy sensors for large-scale environental onitoring purpose. They then used the UAV to discover the topology of the deployed sensor network and repair the network to achieve certain connectivity. Bychkovskiy et al. proposed a sensor network to investigate the control and actuation in data-centric wireless sensor networks (Bychkovskiy, Schoellhaer, & Estrin 1). However, we have observed the following gaps between the proposed active sensor network and the current research work in sensor networks ebedded with robot-driven obility: (1) The robots used in existing research work are either coercial robots which are too expensive to be eployed for real applications, or siple icrorobots with very liited capabilities in coputation, localization and navigation. To develop an active sensor network for real world applications, sensor-network-oriented robots which can be soothly integrated into the underlying sensor network, should be developed. (2) In existing work, no systeatic odels, approaches and ethodologies have been developed regarding the control and anageent of ultiple obile robots in the context of a sensor network. The scalability proble, which certainly needs to be solved in large active sensor networks, has not been addressed yet. This paper is organized as follows: First we present the overall fraework of the active sensor network and the hardware platfor. Then we introduce the sensor node localization proble and propose a distributed, ultiple robot-based localization algorith. Siulation results are provided to verify the algoriths. Overall Fraework In order to provide the above-entioned logistic and network services, we identify the following four core functions that the µservbots should ipleent: (1) sensor localization; (2) service set partition; (3) sensor network-assisted inter-robot counication and (4) distributed task allocation. localization is the process of deterining the geographic location of each sensor, which is very iportant to any sensor network applications. Service set partition ais to divide the sensors into ultiple subsets so that each µservbot is ainly responsible for one subset. This provides a scalable solution to the aintenance of a large network. -network-assisted inter-robot counication utilizes the underlying sensor network to provide a backup counication channel when two robots can not directly talk to each other. Distributed task allocation addresses how to distribute the given task aong ultiple µservbots while axiizing the energy and tie efficiency. Each high level service relies on one or ore of the core functions provided by the µservbots. In this paper, our goal is to ipleent the first two core functions, naely, sensor node localiza- High Level Services Core Functions uservbots Localization Logistic Service Coverage Control Power Supply Calibration All-terrain Mobility Wireless Counication Service Set Partition Network Service Topology Manageent Hierarchical Routing Tie Synchronization Networkassisted Inter-Robot Counication Autonoous Navigation Transportation Energy Harvest/Supply Ranging Distributed Task Allocation Figure 2: The high level services, core functions and capabilities of the µservbots. tion and service set partition. The design of µservbots The first step in developing µservbots is to identify the capabilities needed to ipleent the core functions and the desired logistic and network services. We find the following fundaental capabilities are necessary: all-terrain obility; autonoous navigation; sensor transportation; energy harvest/supply; ranging and wireless counication. Figure 2 shows the hierarchical relationship between the services, the core functions and the capabilities of the µservbots. The conceptual sketch of the µserbot is illustrated in Figure 3. In order to provide all-terrain obility, tracks are adopted for µservbots. Two DC otors provide the driving force in a differential fashion. To provide autonoous navigation capability, the following navigation sensors are ounted on the µservbot: (i) a iniature onidirectional caera; (ii) a iniature GPS; (iii) a 3-axis inclinoeter and (iv) several IR proxiity sensors. The onidirectional caera provides inforation about the local environent surrounding the µservbot and provide visual guidance to approach and attend the sensors. The GPS and the inclinoeter provide location and orientation inforation crucial to the navigation. To provide sensor transportation capability, each µservbot is equipped with a siple gripper. Two servo otors are used to drive the gripper to open/close, and rotate. Energy harvest can take different fors in different applications. Autoatic recharging, vibration energy collection, wind power, etc. are soe exaples. We use onboard solar panels to collect the power for the µservbots. Energy supply to sensors is realized through a novel charge unit on the µservbots. To charge a sensor, a µservbot first picks up a sensor which is ounted in a specially designed universal adaptor (see next section). The gripper then docks the sensor to the onboard charge docking unit, which is ounted on a T-shape supporting frae and consists of a pair of electroagnet-based charging contacts. The polarity of the electroagnets and the charging contacts can be reversed siultaneously to atch the polarity of peranent agnets 479

3 Electroagnetbased (Negative/South) Electroagnetbased (Positive/North) Gripper in Vertical Position Charge Docking Unit Antenna Onidirectional Carea Solar Panels IR sensors Battery Charge Monitoring Charge Docking Electroagnets 3 Axis Inclinoeter PIC18LF67 PIC18LF67 PIC18LF67 GPS Module Gripper Servo Motors Caera Left Motor Xscale (Linux ) PIC18LF67 H-Bridges WiFi I2C Bus Right Motor Acoustic Generator PIC18LF46 Interface Solar Panel Module Rechargeable Battery Gripper in Horizontal Position Figure 3: The conceptual sketch of the µservbot. Plastic Supporting Frae Elastic Holding Ring Bea Peranent Magnetbased (Negative/ South) Peranent Magnetbased (Positive/ North) Plastic Supporting Frae Figure 4: The universal sensor adaptor. Note: all 8 agnetbased contacts are connected to the power contacts on the sensor. and power contacts on the sensor adaptor. This feature can be ipleented by a sall touch sensor ounted against the T-shape supporting frae. If the polarity of the docking unit does not atch the polarity of the sensor adaptor, the generated expulsive force will activate the touch sensor, which in turn will reverse the polarity of the electroagnet and charging contacts through a D flip-flop driven relay. The design of the universal sensor adaptor To enable the robot-sensor interaction, a lightweight, lowcost universal sensor adaptor is developed. Figure 4 shows the design sketch of the universal sensor adaptor. This adaptor consists of a supporting frae and 8 peranent-agnetbased charging contacts with two on each of the four beas. The sensor is held in the two rubber rings attached to the four beas. In order to ake the sensors easy to be recognized by the µservbots, the supporting frae will be painted with certain colors. This universal adaptor serves ultiple purposes: (1) facilitating the battery charging; (2) protecting the sensor by absorbing ipacts during deployent; and (3) aking the sensor easy to identify and grasp. This sensor adaptor can also be equipped with other optional coponents, such as a tiny acoustic generator/receiver to facilitate accurate ranging. The design of µservbot control syste (1) Control architecture. As shown in Figure 5, a hierarchical architecture is adopted for the control syste. At the top Figure 5: The control syste design of the µservbots. level, an Intel Xscale CPU (PXA255 with 64M RAM and 8M flash eory) is responsible for the high level decision aking, otion planning and iage processing. It runs Linux operating syste. At the botto level, PIC icroprocessors are used to control individual sensing and actuation odules. They counicate with the Xscale ain CPU using I 2 C bus. (2) Wireless counication. There are two counication channels on the µservbots. One channel is for direct inter-robot counication, which has higher bandwidth. A Wi-Fi odule (AbiCo Wave2Net Wireless Type I CopactFlash Card) is adopted and driven by the ain CPU. The other channel is for robot-sensor counication, which has lower bandwidth. To facilitate this counication, a sensor interface is designed so that a sensor can be ounted on the µservbot through this interface. Zigbee based MicaZ otes fro Crossbow Technologies Inc.(Xbo ) are used as the sensors. Therefore a µservbot is able to talk to the sensors over Zigbee. (3) Sensing. The sensing odule consists of the onidirectional caera, the GPS (Lassen SQ GPS receiver), the 3-axis inclinoeter (Seika N3) and the set of IR proxiity sensors. Due to the high data rate, the iage processing is ipleented in the ain CPU. The iniature GPS is connected to the ain CPU through serial counication while the 3-axis inclinoeter and the IR sensors are anaged by PIC icroprocessors. (4) Actuation. The actuation odule consists of three units: the ain otor drive unit, the gripper drive unit and the docking control unit. The ain otor drive unit includes a PIC icroprocessor, two H-bridges and two optical encoders for speed feedback. The gripper drive unit includes a PIC icroprocessor, two servo otors and the associated circuit. The docking control unit includes the relay and electroagnets. (5) Power. The power odule consists of two solar panels, a rechargeable Li-Ion battery, and a charging circuit which is responsible for the anageent, onitoring of the charging of the onboard robot batteries and the sensor batteries. The distributed sensor localization algorith An iportant proble in any sensor network applications is to find out the geographic locations of the sensor nodes. In recent years, several sensor localization ethods have been developed for ad hoc wireless sensor networks. Most of the node localization algoriths are based on range easureent, through either tie of arrival (TOA) (Zhao & Guibas 48

4 4), tie difference of arrival (TDOA) (Savarese, Rabaey, & Beutel 1), or received signal strength (RSS) (Bulusu, Heideann, & Estrin ). For exaple, In the Picoradio project (Beutel 1999) at UC Berkeley, a geolocation schee for an indoor environent is provided based on RF received signal strength easureents and pre-calculated signal strength aps. The AHLoS (Ad-Hoc Localization Syste) (Savvides, Han, & Strivastava 1) proposed by Savvides et. al enables sensor nodes to discover their locations using a set distributed iterative algoriths. An RF based proxiity ethod was developed by (Bulusu, Heideann, & Estrin ), in which the location of a node is given as a centroid generated by counting the beacon signals transitted by a set of beacons pre-positioned in a esh pattern. Other ethods that do not rely on range easureents were also developed. For exaple, the count of hops is used as an indication of the distance to the beacon nodes in soe applications (Zhao & Guibas 4). In this paper, we develop a µservbot-assisted distributed node localization algorith. Since each µservbot is equipped with a positioning odule such as the GPS, the µservbots can play as obile beacons in the localization process. The basic idea is as follows. The µservbots ove around and frequently send out localization broadcasts, which are RF essages ebedded with senders current location. A sensor receiving localization broadcasts fro 3 or ore different locations is able to recover its own location through lateration, using the distances calculated fro the strength of the received signals (Bulusu, Heideann, & Estrin ). If the µservbots are equipped with an acoustic generator and the sensors are equipped with acoustic receivers on the adaptor, ore accurate distance easureent can be obtained through Tie Difference of Arrival (TDOA) technique. In order to localize all the sensors, a ethod is needed to ensure that any sensor in the area of interest is able to hear at least 3 different localization broadcasts (RF essage or RF essage with acoustic signal) fro one or ore µservbots. To enure sufficient accuracy, the localization broadcasts should be sent fro distinct locations. On the other hand, in order to save power and iniize the localization tie, we also want to iniize the total traveling distance of the µservbots. To achieve this, the area of interest A, usually a rectangle defined by its four vertices, is placed with grids as shown in Figure 6. The spacing is set to 2 2 r b, where r b is the sensor RF range, or the iniu of the sensor RF range and the acoustic signal range. Therefore, as long as every grid point is visited by one of the µservbots and a localization broadcast is sent out at that grid point, any sensor will receive at least four distinct localization broadcasts. To achieve this, a schee is needed to coordinate the µservbots to visit all the grid points. The proposed cooperative ulti-robot sensor localization will be conducted in the following three steps: (1) µservbots disperse and partition the area of interest into subareas that have roughly equal size. (2) Each µservbot visits the grid points in its associated subarea and sends out localization broadcast at each grid point. (3) Each µservbot collects the location A2 A1 rb A3 Figure 6: Locating the sensors with µservbots. A5 A4 Voronoi Diagra Counication Range inforation of the sensors in its own subarea. In the first step, µservbots disperse theselves into the area of interest A. To achieve unifor dispersion, a potential-based algorith is adopted (Howard, Mataric, & Sukhate 2). The basic idea is to take the boundary of the area of interest A as obstacles that the µservbots should keep away fro, and the µservbots expel each other using potential forces until a stable deployent is achieved. The detailed algorith is as follows: Potential-based deployent algorith for µservbot R i /* starting fro the initial position p i. r ij = p j p i is the relative position fro µservbot R i to µservbot R j. r ik = O k p i is the relative position fro R i to obstacle O k */. (1) Exchange location inforation p i with other µservbots. (2) Calculate the the force caused by the obstacles: 1 F oi = k o r rik 2 ik r ik Calculate the force caused by other robots: 1 F ni = k n r rij 2 ij r ij Calculate the overall force exerted on robot R i: k j F i = F oi + F ni (3) If F i F th then change the velocity of R i according to the following equation: v i = v i + (F i κv i)/ T where F th is a sall force threshold. κ is a viscous friction coefficient. (4) Update the location of R i through the following equation: p i = p i + v i T (5) Go to (1). Here k o and k n are force constants for robot-obstacle interaction and inter-robot interaction, respectively. The settings of these constants will deterine the contribution of each coponent of force in the net force applying on the µservbot. The viscous friction coefficient κ will help iniize oscillations and ensure that the syste will reach steady state as the forces approach zero. T is the tie step for each iteration and is the ass of the µservbot. As proved in (Howard, Mataric, & Sukhate 2), due to the existence of the viscous force κv i, the µservbots will Grid 481

5 Figure 7: The rando distribution of the µservbots and the sensor nodes. The triangles represent the µservbots and the stars represent the sensor nodes. Figure 9: The partition of the area based on Voronoi diagra Figure 8: The trajectories of the µservbots. eventually converge to an equilibriu point, which is the final deployent. To ake the algorith easy to scale up, each µservbot can only interact with its neighbors instead of all the robots, which will aintain the sae convergence due to the fact that reote µservbots generate very sall forces. Once the µservbots are deployed, a Voronoi Diagra (Fortune 1992) D v is constructed, which partitions A into subareas A = {A 1, A 2,..., A n }. It also gives a partition of the grid points set S g into {S g1, S g2,..., S gn }. In the second step, µservbot R i will visit the grid points in its associated set S gi. At each grid point, a localization broadcast will be sent out. Due to the regular pattern of the grid points, the path planning is trivial. For exaple, a zigzag pattern can be used to visit all the grid points. Upon receiving four localization broadcasts, a sensor calculates its location through lateration. In the third step, the location inforation of all the sensors in a subarea is collected by the associated µservbot, which can be done by sending query essages to the corresponding sensors. Siulation Results We conducted siulation of our distributed localization algorith in C and Matlab. A region of 1 1 is con Figure 1: The localization results: circles represent the estiated sensor locations and the stars represent the actual sensor locations. sidered. Five µservbots are randoly deployed. The initial rando locations of the µservbots and the sensor distribution are shown in Fig 7. First, the potential field-based robot deployent algorith is applied to copute an optial location for each µservbot to achieve even coverage of the region. The paraeters used in the algorith are as follows: k o = 48., k n = 8., κ =.5, T =.1s, = 1.kg. The paths followed by the µservbots during the deployent are shown in Fig 8. Then the Voronoi diagra is constructed to partition the region into 5 approxiately equivalent subareas, as shown in Figure 9. The application of the localization algorith in each subarea then follows. Each µservbot oves in a zigzag fashion along the grid points. The RF range, or the iniu of the sensor RF range and the acoustic signal range is set to be r b = 1 2. Therefore the grid spacing is 1. To eulate real world situations, as the sensors collect their data, a Gaussian noise is introduced in the robot location estiation and distance easureent. The Gaussian noises introduced in the robot location estiation and distance easureent have a ean value of and a standard deviation of.5. When a µservbot arrives at each grid point it broadcasts its location to its surrounding sensor nodes. Each sen- 482

6 Table 1: Testing results (Gaussian noise ean=, std. dev.=.2) Test runs Average error () Standard deviation () Table 2: Testing results (Gaussian noise ean=, std. dev.=.5) Test runs Average error () Standard deviation () sor node then collects at least four localization broadcasts as the µservbot reaches within its listening area. Then the location of the sensor node can be calculated using lateration ethod. The coplete localization results of the sensor nodes are shown in Fig 1. We also ran the algoriths on different data sets and recorded the average and standard deviation errors. Two sets of Gaussian noises are used: (1) ean = and standard deviation =.2 (2) ean = and standard deviation =.5. The following tables show the average and standard deviation errors. We find the error in the position estiate (in each of the two test cases with different levels of noise) appears to be higher than the level of noise induced in the syste. The reason is due to the cobined uncertainty of the robot locations and the distance easureents. Conclusions This paper proposes a new sensor network architecture by utilizing ultiple icro obile robots. Such a closed-loop sensor network has any erits that traditional sensor networks do not have. As one of the core functions carried out by the obile robots, node localization is studied and a distributed, ultiple robot-based algorith is proposed. Siulation results prove that the algorith is effective and the sensor nodes can be localized with reasonable accuracy. We expect with ore accurate self-positioning algoriths ipleented on the µservbots, such as vision-assisted localization, the sensor node location accuracy can be further iproved. Our future research will focus on the developent of various algoriths for the other core functions that the µservbots should have. References Akyildiz, I.; Su, W.; Sankarasubraania, Y.; and Cayirci, E. 2. A survey on sensor networks. IEEE Coun. Mag. : Beutel, J Master Thesis, Geolocation in a PicoRadio Environent. UC Berkeley. Bulusu, N.; Heideann, J.; and Estrin, D.. Gpsless low cost outdoor localization for very sall devices. Technical Report -729, Coputer science departent,university of Southern California, Los Angles, CA. Bychkovskiy, V.; Schoellhaer, T.; and Estrin, D. 1. Control and actuation in data-centric wireless sensor networks. In INTELLIGENT DISTRIBUTED AND EMBED- DED SYSTEMS workshop. Corke, P.; Hrahar, S.; Peterson, R.; Rus, D.; Saripalli, S.; and Sukhate, G. 4. Autonoous deployent and repair of a sensor network using an unanned aerial vehicle. In IEEE International Conference on Robotics and Autoation. Fortune Voronoi diagras and delaunay triangulations. In Coputing in Euclidean Geoetry, Edited by Ding-Zhu Du and Frank Hwang, World Scientific, Lecture Notes Series on Coputing Vol. 1. Howard, A.; Mataric, M.; and Sukhate, G. 2. Mobile sensor network deployent using potential fields: A distributed solution to the area coverage proble. In DARS 2. Laraca, A.; Brunette, W.; Koizui, D.; Lease, M.; Sigurdsson, S. B.; Sikorski, K.; Fox, D.; and Borriello, G. 2. Making sensor networks practical with robots. In International Conference on Pervasive Coputing. Mainwaring, A.; Polastre, J.; Szewczyh, R.; Culler, D.; and Anderson, J. 2. Wireless sensor networks for habitat onitoring. In Proceedings of Wireless Network and Applications. Savarese, C.; Rabaey, J.; and Beutel, J. 1. Locationing in distributed ad hoc wireless sensor networks. In Proc. 1 Int l Conf. Acoustics, Speech, and Signal Processing (ICASSP 1), volue 4, 37. Savvides, A.; Han, C. C.; and Strivastava, M. 1. Dynaic fine-grained localization in ad-hoc networks of sensors. In ACM SIGMOBILE. Siic, S. N., and Sastry, S. 3. Distributed environental onitoring using rando sensor networks. In Proceedings of the 2nd International Workshop on Inforation Processing in Networks, Xbow, July, 5. Zhao, F., and Guibas, L. 4. Wireless Networks: An Inforation Processing Approach. Elsevier and Morgan Kaufann Publishers. 483