Distributed Estimation for Motion Coordination in an Unknown …cdcl.umd.edu/papers/jgcd12.pdf ·...
Transcript of Distributed Estimation for Motion Coordination in an Unknown …cdcl.umd.edu/papers/jgcd12.pdf ·...
Distributed Estimation for Motion
Coordination in an Unknown Spatially
Varying Flowfield
Cameron K. Peterson∗ and Derek A. Paley†
University of Maryland, College Park, MD, 20742, USA
I. Introduction
This note addresses the development of decentralized motion coordination algorithms in
the presence of an unknown, spatially varying flowfield. Spatiotemporal flowfields are difficult
to model and may contribute to a signification portion of the vehicle’s inertial velocity. Some
existing algorithms support operation in an unknown uniform flow, though the authors are
not aware of any previous result for cooperative control in an arbitrary, unknown flowfield
that varies smoothly in space. Summers et al. account for constant-velocity wind using
adaptive estimates to drive cooperative vehicles in a loiter circle.1 A uniform flowfield with
an added singular point was mapped by Petrich et al. to improve navigation of shallow-water
underwater autonomous vehicles using only a sparse set of GPS measurements, assuming all
measurements are shared in a centralized fashion.2 Lawrance et al. mapped out a flowfield
using Gaussian process regression for a single vehicle.3 Also for a single vehicle, Langelaan
et al. computed a 3D wind estimate and the wind acceleration using sensors typically found
on a small unmanned aerial vehicle.4
The distributed control algorithm described in this note is achieved using a combined
information and consensus filter.5,6 Lynch et al. first used an information-consensus fil-
ter to estimate spatially varying environmental fields such as temperature.7 A consensus
filter asymptotically convergences to the average of the consensus inputs for either a con-
nected undirected graph5 or a strongly connected and balanced digraph.6 Casbeer and
∗Graduate student, Department of Aerospace Engineering; [email protected]. AIAA Student Mem-ber.†Associate Professor, Department of Aerospace Engineering and Institute for Systems Research;
[email protected]. AIAA Associate Fellow.
1 of 11
Beard showed that the state of a dynamical system that is estimated using an information-
consensus filter is comparable with one obtained from a centralized estimator.8 Olfati-Saber
also provided decentralized Kalman filter formulations,9,10,11,12 developed techniques appli-
cable to a heterogeneous group of sensors,9 and established the properties of stability for the
information-consensus filter.12
In this note, a distributed information-consensus filter is implemented to estimate the
coefficients of a parameterized flowfield. The inter-vehicle communication constraints may
be directed, provided they are strongly connected and balanced.6 (A strongly connected
graph ensures that a communication path that obeys edge direction may be found from
any vehicle to any other vehicle; the communication graph is balanced if, for each vehicle,
the number of incoming connections is equal to the number of outgoing connections.) The
estimated flowfield at the vehicle locations are fed into a decentralized multi-vehicle control
law that cooperatively stabilizes vehicles to a moving formation.
The contribution of this note is a distributed, observer-based control algorithm for the
stabilization of a circular formation in an estimated spatially varying flowfield using a de-
centralized information-consensus filter with noisy position measurements. It is the first
publication that the authors are aware of to consider motion coordination with distributed
estimation of a spatially varying flowfield. In the authors’ previous work,13 the vehicles op-
erated in a spatially uniform flowfield and the estimation scheme was not distributed. In
other previous work,14 the vehicles operated in a flowfield that was entirely known a priori.
The contribution of this note enables operation of multiple vehicles in an arbitrary, random
flowfield assuming only noisy position measurements. Although we present the results for
circular formations of a self-propelled particle model, we expect that the general framework
would apply to other multi-vehicle control systems.
The note proceeds as follows. Section II introduces the vehicle model and outlines an
algorithm for decentralized estimation of a scalar field. Section III proposes a control
algorithm that uses an information-consensus filter to stabilize a circular formation first in
a parametrized flowfield and then in an arbitrary, random flowfield. Section IV provides
concluding remarks.
II. Multi-Vehicle Control and Distributed Estimation
Each vehicle is modeled as a self-propelled Newtonian particle, where the position of
particle k = 1, . . . , N is denoted by rk. The particle travels in a plane at constant, unit speed
relative to an ambient flowfield fk = f(rk) and is subject to a steering control perpendicular
to the velocity eiθk of the vehicle relative to the flowfield. The particle’s inertial velocity is
thus rk = eiθk + fk.14 The steering control uk = θk is the turn rate of the orientation θk
2 of 11
of the velocity relative to the flow. In terms of the inertial speed sk = |rk| and orientation
γk = arg(rk) the equations of motion are
rk = sk(t)eiγk
γk = νk,(1)
where νk is the angular rate of change of the inertial velocity orientation of particle k.
A circular formation is obtained when the instantaneous center of rotation14 ck = rk +
ω−10 ieiγk of each vehicle’s trajectory is fixed and identical. Steering control νk = ω0sk(t)
ensures particle k will traverse a circle with a fixed center ck(0) and a constant radius
|ω0|−1 = |ck(0) − rk(0)|.14 In a circular formation, cj = ck for all pairs j and k and the
vector c = [c1, . . . , cN ]T of circle centers is in the null space of the projection matrix P =
diag{1} − (1/N)11 T , where 1 , (1, . . . , 1)T ∈ RN . Given a known flowfield, one can use
Lyapunov function S(r,γ) , 1/2〈c, Pc〉15 to show that the controla
νk = ω0(sk +K〈Pkc, eiγk〉), K > 0, (2)
stabilizes N vehicles to a fixed circular formation [13, Theorem 1]. The time derivative of S
along solutions of (1) is negative semi-definite with S = 0 occurring when (2) evaluates to
νk = ω0sk(t) and Pc = 0, i.e., the vehicles each travel around a circle, and the circle centers
are identical. (Details of the proof are omitted due to space constraints.) The use of (2)
to stabilize a circular formation in an unknown spatially-varying flowfield is illustrated in
Section III.
In this note, it is assumed that each vehicle can measure or approximate the local flowfield
at its current position. The measurements are incorporated into an information-consensus
filter to estimate the parameters of a global, spatially varying flowfield model. This note
follows Lynch et al.7 in which an information filter and a consensus filter were used to estimate
a scalar environmental field from measurements collected by multiple vehicles. The field fk
at position rk is modeled as7
fk = f(rk) =m∑m=1
n∑n=1
am,nψm,n(rk),
where ψk , ψ(rk) = [ψ1,1(rk), ψ1,2(rk), ..., ψm,n(rk)]T is the vector of l = m× n known basis
functions evaluated at rk and a = [a1,1, a1,2, ..., am,n]T is the set of unknown coefficients to
aPk denotes the kth row of P .
3 of 11
be estimated. This note adopts the following basis vectors:
ψm,n(rk) = cos(πmY
Im(rk))
cos(πnX
Re(rk)), (3)
where m = 1, . . . , m, n = 1, . . . , n, and X and Y represent the dimensions of the domain in
the complex plane.
In this note, it is assumed that the coefficients are constant, i.e., am,n = 0 for all m and
n; they are estimated using either noisy measurements of the flowfield or local approxima-
tions of the flowfield derived from noisy measurements of vehicle position. Each flowfield
measurement fk is corrupted by Gaussian, zero-mean measurement noise vk with variance
Rk ∈ C, so that7
fk = ψTka+ vk. (4)
The information filter is a variation of the Kalman filter that propagates forward the
inverse of the error covariance.16 Let a be the estimated flowfield coefficients and M =
E[(a−a)(a−a)T ] be the coefficient error covariance.b The inverse error covariance M−1 , I
is called the information matrix and i = Ia is the information measurement.7 This note
implements a discrete form of the information filter. Let t be the current time and ∆t
the time step. The superscript (−) denotes the prior estimates and (+) denotes the updated
estimates. The information filter equations are simplified under the assumption that the state
a is constant and does not have process noise. These conditions imply that the predicted
information covariance and information state at time t are equal to the prior values, i.e.,
I−(t) = I+(t−∆t) and i−(t) = i+(t−∆t). The measurement update equations for particle
k are16,7
I+k = I−k +ψkR
−1k ψ
Tk
i+k = i−k +ψkR−1k fk.
Rewriting these equations using Ck , ψkR−1k ψ
Tk and yk , ψkR
−1k fk yields7
I+k = I−k + Ck
i+k = i−k + yk.(5)
The matrix Ck and vector yk represent the information gained from particle k in a single
update measurement. The estimated coefficients ak for particle k are obtained from the
information matrix using ak = I−1k ik.
7 An advantage of using the information filter is that
measurement updates are simply added to the predicted information matrix and predicted
information measurement.8 Measurements from N vehicles can be incorporated in a single
bE[·] is the expected value of [·].
4 of 11
update step using the following sums:7
C ,N∑k=1
Ck =N∑k=1
ψkR−1k ψ
Tk (6)
and
y ,N∑k=1
yk =N∑k=1
ψkR−1k fk. (7)
The measurement-update equations that incorporate the information from all particles are
I+ = I− + C
i+ = i− + y,(8)
with the estimated coefficients a = I−1i.
A centralized information filter can be used directly to estimate a when all-to-all commu-
nication is available. When all-to-all communication is not available, the information filter
is supplemented by a consensus filter. A consensus filter approximates the average value
of an input parameter and converges to the true average as long as the (directed) vehicle
communication topology is strongly connected and balanced.6 The information-consensus
filter allows vehicle k to approximate C and y using information from its neighbor set Nk,where j ∈ Nk indicates that vehicle k receives communication from vehicle j. Let C(i,j),k
indicate the entry in the ith row and jth column of Ck. Likewise yn,k is the nth entry of
vehicle k’s measurement vector. Let τ0,k be particle k’s input to the estimated value. That is,
τ0,k = C(i,j),k where i, j = 1, . . . , l or τ0,k = yn,k where n = 1, . . . , l. The proportional-integral
(PI) consensus filter is7
τk = ξ(τ0,k − τk)−KP
∑j∈Nk
(τk − τj) +KI
∑j∈Nk
(ηk − ηj)ηk = −KI
∑j∈Nk
(τk − τj).(9)
The gain ξ > 0 determines how much the consensus filter relies upon its own input relative to
the inputs from other connected particles. τk is the consensus variable, i.e., the approximate
average of C(i,j),k or yn,k, ηk is an integrator variable, and KP and KI are the proportional
and integral gains, respectively. The sums in (9) are computed for all the particles in the
neighbor set of k.
The next section provides a distributed motion coordination algorithm that estimates a
spatially varying flowfield using (8) with inputs C and y approximated by the consensus
filter (9).
5 of 11
Nf~
.
.
.
IF
Control
Nν
2ν...
2~f
1~f 1ν
1r
2r
Nr
+
NN yC ,
22 , yC
11, yC
yC, f
(a) Centralized architecture
.
.
.
Control
Nν
2ν...
1~f 1ν
1m
2m
Nm
11, yC 11, yNCN
Nf
.
.
.
.
.
.
CF
CF
CF
IF
IF
IF
2f
1f
22 , yNCN
NN yNCN ,
22 , yC
NN yC ,
2~f
Nf~
(b) Decentralized architecture
Figure 1. Flowfield estimation and multivehicle control architectures.
III. Stabilization of Circular Formations in an Estimated
Spatially Varying Flowfield
Flowfield estimation and control may be accomplished in a centralized manner using
an information filter or in a decentralized manner using an information-consensus filter.
Figure 1 illustrates both designs. In Figure 1(a), the flowfield fk is approximated by a set
of basis functions as in (4), where the basis vector is known and the flowfield coefficients
are estimated. At each time step, vehicle k measures the local flowfield at its position rk.
Equations (6) and (7) are used to obtain C and y, which represent the information gained
from the local flowfield measurement. The centralized filter uses C and y to compute the
global flowfield estimate f . The estimate f (and its directional derivative) are fed into each
particle’s steering controller. This process is repeated at each time step and the global
flowfield estimate improves from the additional measurements.
The decentralized algorithm, depicted in Figure 1(b), uses a PI consensus filter (9) to
calculate Ck, the approximate average of matrix (6), and yk, the approximate average of
measurement vector (7). This note also relaxes the assumption that each particle measures
the local flowfield and instead requires only noisy position measurements as described next.
Let mk(t) be the change in vehicle position from t−∆t to t, subject to measurement error
gk(t):
mk(t) = rk(t) + gk(t)− rk(t−∆t)− gk(t−∆t). (10)
Using rk(t) = lim∆t→∞(rk(t)− rk(t−∆t))/∆t in (10) yields
mk(t) ≈ rk∆t+ gk(t)− gk(t−∆t)
≈ [eiθk(t) + fk(t)]∆t+ gk(t)− gk(t−∆t).
For a sufficiently small ∆t, one can assume θk is constant without loss of generality. The
6 of 11
Table 1. Decentralized Information-Consensus Filter Cooperative Control Algorithm
Input: Basis vector ψ, sensor variances Rk, circle formation radius |ω0|−1, and a stronglyconnected and balanced communication topology
At each time step t, each particle k = 1, . . . , N performs the following steps:
1: Take a noisy position measurement rk
2: Use the difference between the previous and current position measurement to approxi-mate the local flowfield measurement using (11)
3: Evaluate the basis vector at the measured position: ψk , ψ(rk)
4: For n = 1, . . . , p, where p is the number of consensus filter iterations, repeat:
4a: Use the consensus filter to estimate the components of C and y
5: Update the approximate prior information matrix I− and measurement i− using (8)and determine the estimated coefficients ak = I−1i
6: Compute the estimated flowfield fk = ψkak = ψkI−1i
7: Compute control νk using (2) with sk, c and γk replaced by their estimated values
approximate local flowfield measurement fk(t) at time t is thus
fk(t) = mk(t)∆t− eiθk(t) + gk(t)−gk(t−∆t)
∆t. (11)
Note time step ∆t must be small enough so that θk can be considered constant, but not so
small that the change in position measurement error from t − ∆ to t dominates (11). In
order to approximate a local flowfield each vehicle also needs to know the orientation θk of
its velocity relative to the flow. (The speed relative to the flow is one by assumption.)
Communicating only with vehicles in its neighbor set, each particle uses a consensus
filter to determine an approximate average of Ck and yk. These values are multiplied by
the number of particles N to approximate C and y, which are used by the information
filters to generate estimates of the flowfield coefficients. Each particle may have different
coefficient estimates due to variances in the approximate average of the covariance and
measurement matrix. The estimated flowfield is used in each vehicle’s steering control and
the process is repeated at the next time step. Table 1 and the following proposition present
the information-consensus filter algorithm.
Proposition 1. Let f(rk) =∑m
m=1
∑nn=1 am,nψm,n(rk) be a spatially varying flowfield where
the ψm,n(rk) are known basis vectors and the coefficients am,n are unknown. Using the algo-
rithm described in Table 1 forces convergence of solutions of model (1) to the set of a circular
7 of 11
formations with radius |ω0|−1 and direction of rotation determined by the sign of ω0.
Proposition 1 is justified because the estimation stage of the algorithm in Table 1 (steps
3–6) is separate from the control dynamics. The combined information-consensus filter
behaves like the centralized information filter, provided it is given time to converge to the
average. Convergence of the consensus filter is assured for strongly connected and balanced
communication topologies. Due to the stability properties of the information filter12 the
flowfield estimate will improve independently from the steering control. When the time step
is sufficiently small to ensure that the local flowfield is approximately uniform, then (11)
is an adequate replacement for the noisy flowfield measurement of Table 1. The control
becomes more accurate as the estimated flowfield converges to the true flowfield. Once the
flowfield is estimated, as discussed in Section II, steering control (2) drives all particles to a
circular formation with identical center points and radii [13, Theorem 1] .
Figure 2 illustrates the results of simulating the decentralized information-consensus algo-
rithm using noisy position measurements in flowfield (3) with l = 10 known basis functions.
Figure 2(a) depicts the vehicles (red circles), the vehicle trajectories (blue tracks), and the
inertial velocities (black arrows) at t = 500 seconds. The actual flowfield is depicted by the
gray vector field in the background. The vehicles were commanded to a circular formation
with radius |ω0|−1 = 10. Each particle approximates the local flowfield using (11) and uses
that approximation in an information-consensus filter to estimate the global flowfield. The
particles have a strongly connected and balanced topology, communicating with only four
neighbors, such that particle k receives communication directly from particles k − 2, k − 1,
k + 1 and k + 2, modulus N . We set KI = 0.05, KP = 0.5, ξ = 0.01, and sensor variance
Rk = 0.01. Figure 2(b) shows the error in each basis coefficient estimate converging to zero
for particle k = 15. The root mean square error (RMSE) of the coefficient estimates is
depicted by the black line. The error values for the decentralized consensus filter take longer
to converge than the centralized implementation, because the imperfect estimates of C and
y increase the duration of the transient.
Figure 3 illustrates the results of simulating the decentralized information-consensus al-
gorithm using noisy position measurements in an arbitrary, random flowfield modeled by
(but not necessary spanned by) l = 10 basis functions of the form (3). Figure 3(a) depicts
the particles as they are converging to a circular formation with the actual (gray arrows)
and estimated (gold arrows) flowfield as background vectors. Figure 3(c) shows the particles
at time t = 500 seconds, after they have achieved a circular formation. Figures 3(b) and
3(d) depict the percent flowfield error between the actual and estimated flowfield for particle
k = 15 at times t = 100 and t = 500 seconds. As time progresses, the estimate of the global
flowfield improves, although the estimation errors are higher toward the edges of the region
where fewer measurements were collected.
8 of 11
0 10 20 30 40 500
10
20
30
40
50
Re(r)
Im(r
)
(a) t =[0, 500] s.
0 50 100 1500
0.5
1
1.5
Time (s)
Coe
ffici
ent E
rror
(b) Flowfield coefficient errors, k = 15.
Figure 2. Stabilization of a circular formation in an estimated spatially varying flowfield usingan information-consensus filter driven by noisy position measurements.
IV. Conclusion
This note describes the design of a decentralized control algorithm for autonomous ve-
hicles that operate in an estimated spatially varying flowfield. The authors provide a dis-
tributed motion coordination algorithm that estimates a spatially varying flowfield and uses
the estimate in a closed-loop multi-vehicle control. Each vehicle implements an information-
consensus filter to reconstruct the flowfield parameters from noisy position measurements.
Simulations are provided to illustrate the performance of the estimated control algorithms in
stabilizing a circular formation in a parametrized spatially variable flowfield. This approach
is shown to be viable in an arbitrary, random flowfield.
Acknowledgments
This material is based upon work supported by the National Science Foundation under
Grant No. CMMI0928416 and the Office of Naval Research under Grant No. N00014-09-1-
1058. The authors would like to thank Sharan Majumdar, Kayo Ide, and Pat Murphy for
discussions related to this note and the comments of the reviewers.
References
1Summers, T. H., Akella, M. R., and Mears, M. J., “Coordinated Standoff Tracking of Moving Targets:
Control Laws and Information Architectures,” Journal Guidance, Control, and Dynamics, Vol. 32, No. 1,
2009, pp. 56–69.
9 of 11
0 10 20 30 40 500
10
20
30
40
50
Re(r)
Im(r
)
(a) t =[0, 100] s.
Re(r)
Im(r
)
0 10 20 30 40 500
10
20
30
40
50
0
10
20
30
40
(b) Flowfield percent errors, t = 100 s.
0 10 20 30 40 500
10
20
30
40
50
Re(r)
Im(r
)
(c) t =[0, 500] s.
Re(r)
Im(r
)
0 10 20 30 40 500
10
20
30
40
50
0
10
20
30
40
(d) Flowfield percent errors, t = 500 s.
Figure 3. Stabilization of a circular formation in an estimated spatially varying flowfield usinga decentralized information-consensus filter in a randomly generated flowfield.
2Petrich, J., Woolsey, C. A., and Stilwell, D. J., “Planar Flow Model Identification for Improved
Navigation of Small AUVs,” Ocean Engineering , Vol. 36, No. 1, Jan. 2009, pp. 119–131.
3Lawrance, N. R. J. and Sukkarieh, S., “Autonomous Exploration of a Wind Field with a Gliding
Aircraft,” Journal of Guidance, Control, and Dynamics, Vol. 34, No. 3, 2011, pp. 719–733.
4Langelaan, J. W., Alley, N., and Neidhoefer, J., “Wind Field Estimation for Small Unmanned Aerial
Vehicles,” Journal of Guidance, Control, and Dynamics, Vol. 34, No. 4, 2011, pp. 1016–1030.
5Olfati-Saber, R., Fax, J. A., and Murray, R. M., “Consensus and Cooperation in Networked Multi-
Agent Systems,” Proceedings of the IEEE , Vol. 95, No. 1, 2007, pp. 215–233.
6Ren, W., Beard, R. W., and Atkins, E. M. A., “A Survey of Consensus Problems in Multi-Agent
Coordination,” American Control Conference, 2005. Proceedings of the 2005 , IEEE, Portland, Oregon, June
2005, pp. 1859–1864.
7Lynch, K. M., Schwartz, I. B., Yang, P., and Freeman, R. A., “Decentralized Environmental Modeling
by Mobile Sensor Networks,” IEEE Transactions on Robotics, Vol. 24, No. 3, 2008, pp. 710–724.
10 of 11
8Casbeer, D. W. and Beard, R. W., “Distributed Information Filtering using Consensus Filters,” Amer-
ican Control Conference, St. Louis, Missouri, 2009, pp. 1882–1887.
9Olfati-Saber, R., “Distributed Tracking for Mobile Sensor Networks with Information-Driven Mobil-
ity,” American Control Conference, 2007, pp. 4606–4612.
10Olfati-Saber, R. and Shamma, J. S., “Consensus Filters for Sensor Networks and Distributed Sensor
Fusion,” IEEE Conference on Decision and Control , Seville, Spain, Dec. 2005, pp. 6698–6703.
11Olfati-Saber, R., “Ultrafast Consensus in Small-World Networks,” American Control Conference, Port-
land, Oregon, June 2005, pp. 2371–2378.
12Olfati-Saber, R., “Kalman-Consensus Filter: Optimality, Stability, and Performance,” IEEE Confer-
ence on Decision and Control held jointly with Chinese Control Conference, Dec. 2009, pp. 7036–7042.
13Peterson, C. K. and Paley, D. A., “Multivehicle Coordination in an Estimated Time-Varying Flow-
field,” Journal of Guidance, Control, and Dynamics, Vol. 34, No. 1, Jan. 2011, pp. 177–191.
14Paley, D. A. and Peterson, C., “Stabilization of Collective Motion in a Time-Invariant Flow Field,”
Journal of Guidance, Control, and Dynamics, Vol. 32, No. 3, 2009, pp. 771–779.
15Sepulchre, R., Paley, D. A., and Leonard, N. E., “Stabilization of Planar Collective Motion: All-to-All
Communication,” IEEE Transactions on Automatic Control , Vol. 52, No. 5, 2007, pp. 811–824.
16Simon, D., Optimal State Estimation: Kalman, H [infinity] and nonlinear approaches, John Wiley &
Sons, Inc., New Jersey, 2006.
11 of 11