Kostas Compass3 30

Embed Size (px)

Citation preview

  • 7/30/2019 Kostas Compass3 30

    1/35

    Robotics and Sensor Networks:

    Coverage, Localization and Mobility

    Kostas Bekris

    March 29, 2005COMPASS project meeting

  • 7/30/2019 Kostas Compass3 30

    2/35

    What is the relation?

    Robotics and Sensor Networks are typically

    considered two unrelated fields.

    But: Robots can provide mobility to Sensor Networks.

    Sensor Networks can provide rich sensing

    information to Robots.

    and most importantly

    The two fields are facing many similar challenges.

  • 7/30/2019 Kostas Compass3 30

    3/35

    Robots for Sensor Networks

    Mobile nodes can be used to:

    Re-deploy and calibrate sensors,

    React to sensor failures and

    Deliver power.

    [Corke, Hrabar, Peterson, Rus, Saripalli, Sukhatme, 2004]

  • 7/30/2019 Kostas Compass3 30

    4/35

    Sensor Networks for Robots

    A network offers detailed

    sensing information to a

    robot that is not possible

    to acquire otherwise.

    Distributed computation

    over the network.

    Robots can form mobile

    sensor networks.[Batalin, Sukhatme, Hattig, 2004]

  • 7/30/2019 Kostas Compass3 30

    5/35

    Similar challenges

    Many of the problems are the same:

    Decision inference based on multiple sensing inputs

    Sensor fusion

    Location awareness Coordination

    Task allocation

    Workspace or sensor field coverage

    Compression of data

    Uncertainty

    Mobility

  • 7/30/2019 Kostas Compass3 30

    6/35

    Topics to cover

    I. CoverageArt-Gallery Problems

    (Computational Geometry)

    II. LocalizationDistributed Markov and Monte Carlo

    (Machine Learning)

    III. MobilityArtificial Potential Functions &

    Formation Control

    (Control Theory)

  • 7/30/2019 Kostas Compass3 30

    7/35

    I. Coverage

  • 7/30/2019 Kostas Compass3 30

    8/35

    Coverage in Sensor Networks

    Very important for deployment:

    Under-deployment might result in communication

    failures or failures in the sensing task

    Over-deployment can significantly increase the cost

    Typical Measure in

    Sensor Networks:

    Path Exposure

    [Meguerdichian, Koushanfar, Potkonjiak, Srivastava 2001]

  • 7/30/2019 Kostas Compass3 30

    9/35

    Art-Gallery Problem

    The original art gallery problem:

    Find the smallest number of point

    guards g(n) necessary to cover any

    polygon ofn vertices.

    According to the art gallery theorem the necessary

    number is: g(n) = n/3

    Finding minimum set of guards: NP-hard

    [Conversation between Klee and Chvatal 1972]

    [Chvatal 1975]

    [Aggarwal 1984]

  • 7/30/2019 Kostas Compass3 30

    10/35

    Heuristic Solution

    Greedy approach for map building in robotics:

    Place the first guard at the point of

    maximum visibility Next guard is placed where it sees the maximum

    area not visible to the first and so on

    The sub-problem of finding the next guard of

    maximum visibility is called:

    the Next-Best-View problem

  • 7/30/2019 Kostas Compass3 30

    11/35

    Various approaches

    Randomized algorithms compute the optimal

    location up to a constant factor approximation.

    Sampling-based techniques can be used for themost realistic case of sensors with limited-range.

    Decomposition methods

    compute cells that can be

    observed by limited range

    guards.

    [Cheong, Efrat, Har-Peled 2004]

    [Kazazakis, Argyros 2002]

    [Gonzalez-Banos, Latombe 2002]

  • 7/30/2019 Kostas Compass3 30

    12/35

    Robotic SN Deployment

    [Howard, Mataric, Sukhatme 2002]

    Incremental approach: select a node at a time to be

    deployed in a new location, a second nodes replaces it

    Build a centralizedrepresentation

    while maximizing

    network coverage

    and retaining

    line-of-sight

    communication.

  • 7/30/2019 Kostas Compass3 30

    13/35

    II. Localization

  • 7/30/2019 Kostas Compass3 30

    14/35

    Data for SN self-localization

    Received Signal Strength: for known transmission

    power, the propagation loss is measured to estimate

    the distance based on a propagation model.

    Time-of-arrival or time-difference-of-arrival: The

    propagation time can be directly translated into

    distance based on signal propagation speed.

    Angle-of-arrival: Systems estimate the angle at

    which signals are received.

  • 7/30/2019 Kostas Compass3 30

    15/35

    Localization Approaches

    [Bergamo,

    Mazzini 2002]

    Assume a subset of the nodes can self-localize(e.g. GPS) localize the rest relative to the beacons.

    Trilateration Triangulation MLE

    [Niculescu,

    Nath 2003][Nasipuri, Li 2002]

    [Savvides, Han,

    Srivastava 2002]

  • 7/30/2019 Kostas Compass3 30

    16/35

    Uncertainty in Robotics

    [Fox, Burgard, Kruppa, Thrun: A probabilistic approach

    to collaborative multi-robot localization, 2000]

    Robots, like nodes of sensor networks, have to be

    aware of their location.

    Typical sensors in robotics: sonar, laser, cameras.

    Problem: inherent uncertainty in sensor measurements

    Probabilistic/bayesian techniques proven successful

    in dealing with uncertainty and providing robustness.

  • 7/30/2019 Kostas Compass3 30

    17/35

    Markov Localization

    Each robot maintains a belief for its position at time t

    Belt(L)

    where L is the robots configuration (e.g. {x,y, }).

    Initially, Bel0(L) follows a uniform distribution.

    Each robot collects data dt:

    (a) Odometry: at

    (b) Sensing observations: ot

    (c) Detections of other robots: rt

  • 7/30/2019 Kostas Compass3 30

    18/35

    Updating the distribution

    The belief represents the posterior up to time t:

    Belt(L) = Pr(Lt|dt)

    Perception model:

    Pr(ot|L)

    Motion Model:

    Pr(L|at,L)

    Updates after:

    (1) Sensing: Belt(L) = Pr(ot|L) Belt-1(L)

    (2) Action: Belt

    (L) = Pr(L|at

    ,L) Belt-1

    (L) dL

  • 7/30/2019 Kostas Compass3 30

    19/35

    Multi-Robot Case

    Independence assumption:

    Pr(L1, , Ln|dt) = Pr(L1|d

    t) Pr(Ln|dt)

    Detections used to add additional constraints.

    Assume robot m detects robot n:

    Beltn(L) =

    Belt-1n(L) Pr(Ln=L|rtm,Lm=L) Bel

    t-1m(L) dL

  • 7/30/2019 Kostas Compass3 30

    20/35

    Monte-Carlo Localization

    Representation issue with the storage of distributions

    Monte Carlo approach:

    A distribution is a set ofKweighted particles:

    S = { (Li,pi) | i=(1,,K) }

    where: Li is a candidate position and

    pi is a discrete probability pi=1

    Sensing leads to re-weighting the set of samples so

    as to agree with the measurements.

  • 7/30/2019 Kostas Compass3 30

    21/35

    An equivalent approach is to distribute thecomputation of a centralized Kalman filter to

    separate Kalman filters.

    More difficult problem: SLAM (Simultaneous

    Localization and Mapping)

    Incrementally generate a maximum likelihood

    map

    Probabilistically estimate the robots position

    More on Localization

    [Roumeliotis, Bekey 2002]

  • 7/30/2019 Kostas Compass3 30

    22/35

    Providing location aware services in buildings thatare equipped with wireless infrastructure

    Build radio signal strength maps with multiple robots:

    For a pair of locations return the expected

    signal strength

    Sample the environment and build the map for the

    samples

    Localization for RSN

    [Hsieh, Kumar, Taylor 2004]

    [Ladd, Bekris, Rudys, Marceau, Kavraki, Wallach 2002]

    [Haeberlen, Flannery, Ladd, Rudys, Wallach, Kavraki 2004]

  • 7/30/2019 Kostas Compass3 30

    23/35

    III. Mobility

  • 7/30/2019 Kostas Compass3 30

    24/35

    Why mobility?

    Synoptic sensing implies either over-deployment(impractical you cannot have sensor everywhere)

    or mobility

    Mobility allows the system to focus sensing where

    it is needed, when it is needed

    The initial deployment of static nodes cannot deal

    with all possible changes in the environment

  • 7/30/2019 Kostas Compass3 30

    25/35

    Energy Considerations

    [Dantu, Rahimi, Shah, Babel,

    Dhariwal, Sukhatme 2004]

    Example Mobile Platform: Robomote

  • 7/30/2019 Kostas Compass3 30

    26/35

    Goal of navigation approaches

    Navigational strategies for SN should not haveextensive sensing and computational requirements.

    They should take advantage of the distributed natureof such networks.

    Computationally or memory expensive approachesare also not appropriate.

  • 7/30/2019 Kostas Compass3 30

    27/35

    Navigation Functions

    Many distributed navigation approaches are basedon navigation functions.

    Construct a real-valued map: V:C

    f R

    with uniqueminimum at the goal and is maximal over Cf boundary.

    [Rimon, Koditschek 1992]

  • 7/30/2019 Kostas Compass3 30

    28/35

    Navigation Functions

    Then the robot at position p can move according to:

    where d is an arbitrary dissipative vector-field.

    Under additional requirements NFs guide the robot to

    the goal without hitting local minima.

    In the multi-robot case, each robot can act as an

    obstacle in the potential function of other robots.

    (p,p) = -V(p) + d(p,p)

    [Dimarogonas, Zavlanos, Loizou, Kyriakopoulos 2003]

  • 7/30/2019 Kostas Compass3 30

    29/35

    Source Gradient Climbing

    A mechanism in the environment may be inducingan environmental gradient field (light, sound source).

    APFs are used for locating the source with multiple

    robots.

    If a robot measures the gradient only in the direction of

    motion then it can only find minima along a line.

    An APF enforces the team to stay close and eventually

    the source will be found. [Ogren, Fiorelli, Leonard 2004]

  • 7/30/2019 Kostas Compass3 30

    30/35

    Formation Control

    Another possibly desirable behavior with a team ofmobile systems is to move the entire team in formation.

    Alternatives such as (l- ) or (l-l) control have been

    considered as basic motion primitives for formations.[Desai, Ostrowski, Kumar 2001]

  • 7/30/2019 Kostas Compass3 30

    31/35

    Conclusion

  • 7/30/2019 Kostas Compass3 30

    32/35

    Our interest

    Interested in networks that have the ability to adaptthe location of their nodes

    - not necessarily with autonomous mobility

    to solve problems that might require node relocation

    Do not assume mobility is easily available and

    inexpensive as it is typically considered in robotics

    Take into account the cost of mobility and apply it only

    when it is necessary for the application

  • 7/30/2019 Kostas Compass3 30

    33/35

    Sampling-Based Motion Planners

    An improvement over potential functions in typicalrobotic applications.

    They sample the configuration space of robots and

    construct lower-dimensional representations

    (e.g. graph structures).

    They solve path planning problems on the graph

    structures.

  • 7/30/2019 Kostas Compass3 30

    34/35

    Issues to consider

    Can we apply the SBMP framework to deal withadaptive sensor network problems?

    Can we have distributed SBMP?

    Can SBMP plan not just for motion but for other tasks,

    such as sensing and communication?

    Can we take into consideration the fact that different

    tasks have different energy costs?

  • 7/30/2019 Kostas Compass3 30

    35/35

    Questions??

    THE END