Controllability.

Example

Motion planning for a vehicle:

Linear Systems η θ

Lecture 3. Controllability P (ξ, η )

φ

Torkel Glad
ξ
Reglerteknik, ISY, Linköpings Universitet
u1 : turning speed of front wheels, u2 : forward speed
θ̇ = u1 , φ̇ = u2 sin θ
ξ̇ = u2 cos(θ + φ), η̇ = u2 sin(θ + φ)

AUTOMATIC CONTROL AUTOMATIC CONTROL

Torkel Glad REGLERTEKNIK Torkel Glad REGLERTEKNIK
Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET

Example, cont’d New directions from Lie brackets

Start at x = xo .
Move along f1 (x), then f2 (x), then −f1 (x), then −f2 (x)
In vector form Move for h units of time along each direction.
 
θ
 
1

0
 The resulting movement is h2 [f1 , f2 ](xo ) + O(h3 )
d  • [f1 , f2 ] = f2,x f1 − f1,x f2 is the Lie bracket.
φ = u1 0 + u2  sin θ 
    
dt ξ
   0   cos(θ + φ) By doing nested movements one can generate [f3 , [f1 , f2 ]],
η 0 sin(θ + φ) [[f1 , f2 ], [f3 , f4 ]], ........
If the set of all possible Lie brackets spans the space, then
How can you span a 4-dimensional space with 2 vectors?
intuitively one should have full controllability.
Problem: for systems of the form
ẋ = f (x) + u1 g1 (x) + · · · + um gm (x) it is only possible to move
along f (x) in the positive direction.

AUTOMATIC CONTROL AUTOMATIC CONTROL

Torkel Glad REGLERTEKNIK Torkel Glad REGLERTEKNIK
Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET
Lie brackets for car example in Matlab Properties of Lie bracket

>> syms th phi x y

>> f1=[1;0;0;0]
>> f2=[0;sin(th);cos(th+phi);sin(th+phi)] [f , g] = gx f − fx g
>> vars=[th,phi,x,y]
Some Lie bracket formulas:
>> f1f2=jacobian(f2,vars)*f1-jacobian(f1,vars)*f2
[a, a] = 0
f1f2 =
[a, b] = −[b, a]
[ 0] [a + b, c] = [a, c] + [b, c]
[ cos(th)] [a, [b, c]] + [b, [c, a]] + [c, [a, b]] = 0 Jacobi identity
[ -sin(th+phi)]
[ cos(th+phi)]

AUTOMATIC CONTROL AUTOMATIC CONTROL

Torkel Glad REGLERTEKNIK Torkel Glad REGLERTEKNIK
Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET

Controllability, definitions Controllability, the test:

AU (x0 ): the reachable set from x0 , while remaining in the set U

The system controllable: ARn (x) = Rn for any x.
ẋ = f (x, u)

Problem: If U is a small neighborhood AU is often “one-sided”

Often the case for systems with drift term: 1. form fj (x) = f (x, uj ), for all possible constant uj
ẋ = f (x) + u1 g1 (x) + · · · + um gm (x) 2. form all possible linear combinations of all possible iterated Lie
brackets of the fj
The natural local property is local accessibility:
3. Hermann and Krener 1977: If they span the state space at x0
The system locally accessible at x:
(controllability rank condition), then the system is locally
• AU (x) has a nonempty interior for any neighborhood U of x
• i.e. AU (x) has full dimension. accessible at x0 .

AUTOMATIC CONTROL AUTOMATIC CONTROL

Torkel Glad REGLERTEKNIK Torkel Glad REGLERTEKNIK
Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET
Controllability, linear systems To reach with minimum energy

Control from x0 to xf with minimum energy:

ẋ(t) = A(t)x(t) + B(t)u(t), x(t0 ) = x0 , x(tf ) = xf

1 t1 T
Z
ẋ = Ax + Bu minimize J = u udt
2 t0

[Ax, bi ] = −Abi , where bi is a column of B The control then has the form
[B, AB, . . . , An−1 B] full rank ⇒ controllability rank condition u(t) = −BT (t)ΦT (t0 , t)W (t0 , t1 )−1 z, z = x0 − Φ(t0 , t1 )xf
satisfied.
The linear structure permits much stronger conclusions. where W is the controllability Gramian
Z t1
W ( t0 , t1 ) = Φ(t0 , t)B(t)BT (t)ΦT (t0 , t) dt
t0

AUTOMATIC CONTROL AUTOMATIC CONTROL

Torkel Glad REGLERTEKNIK Torkel Glad REGLERTEKNIK
Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET

Controllability Controllability, time invariant case

The controllability Gramian

Z t1
W ( t0 , t1 ) = Φ(t0 , t)B(t)BT (t)ΦT (t0 , t) dt
t0 Time-invariant case (A, B constant):

Theorem Possible to move the state from x0 at t = t0 to x1 at t = t1 C = B AB · · · An−1 B

⇔
z = x0 − Φ(t0 , t1 )x1 is in the range space of W (t0 , t1 ). Theorem The range and null spaces of W (t0 , t) coincide with the
range and null spaces of CCT for all t > t0 .
The energy required is

1 T
J= z W ( t0 , t1 ) − 1 z
2

AUTOMATIC CONTROL AUTOMATIC CONTROL

Torkel Glad REGLERTEKNIK Torkel Glad REGLERTEKNIK
Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET
Discrete time Discrete time, time invariant case

Iterating the state equation gives

R(t0 ,t1 )U = z

where R(t0 ,t1 ) attains maximum rank after at most n time steps
(Cayley-Hamilton). The controllability properties are thus given by
U = [uT (t1 −1) . . . uT (t0 )]T
C = B AB · · · An−1 B = R(t0 ,t0 +n)


z = x(t1 ) − A(t1 −1) · · · A(t0 )x(t0 )
R(t0 ,t1 ) = [B(t1 −1) A(t1 −1)B(t1 −2) . . . Controllability theory for time invariant systems is thus the same in
... A(t1 −1) · · · A(t0 +1)B(t0 )] continuous and discrete time (ẋ replaced by x(t + 1)).

It is possible to move from x(t0 ) to x(t1 ) in the time interval [t0 , t1 ] if

and only if z is in the range space of R(t0 ,t1 ).
The minimum value of UT U is zT (RRT )−1 z. RRT thus corresponds
to the continuous time controllability Gramian.
AUTOMATIC CONTROL AUTOMATIC CONTROL
Torkel Glad REGLERTEKNIK Torkel Glad REGLERTEKNIK
Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET

Change of state variables PBH tests

ẋ = Ax + Bu, y = Cx, x = Tx̄, T nonsingular
gives

Theorem A, B controllable if and only if

x̄˙ = Āx̄ + B̄u, y = C̄x
Ā = T −1 AT, B̄ = T −1 B, C̄ = CT, C̄ = T −1 C pT A = λpT , pT B = 0 ⇒ p = 0

Controllability is thus preserved under this similarity transformation. Theorem A, B controllable if and only if
Theorem Let the rank of C be r. Then T can be chosen so that

    rank sI − A B = n
Ā11 Ā12 B̄11
Ā = , B̄ =
0 Ā22 0 for all complex s.

where Ā11 is r × r, B̄11 is r × m and Ā11 , B̄11 controllable.

Note: The partition of eigenvalues between Ā11 and Ā22 is unique:
“controllable” and “uncontrollable” eigenvalues
AUTOMATIC CONTROL AUTOMATIC CONTROL
Torkel Glad REGLERTEKNIK Torkel Glad REGLERTEKNIK
Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET Linear Systems 2010, Lecture 3 LINKÖPINGS UNIVERSITET