C4 BMobile Robots
C4 BMobile Robots
2.1 Holonomicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2
3
4.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.1.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3
4
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.3 Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
9 In Conclusion 89
10 Miscellaneous Matters 90
11 Example Code 94
5
Topic 1
This set of lectures is about navigating mobile platforms or robots. This is a huge topic and
in eight lectures we can only hope to undertake a brief survey. The course is an extension
of the B4 estimation course covering topics such as linear and non-linear Kalman Filtering.
The estimation part of the lectures is applicable to many areas of engineering not just mobile
robotics. However I hope that couching the material in a robotics scenario will make the
material compelling and interesting to you.
Lets begin by dispelling some myths. For the most parts when we talk about “mobile
robots” we are not talking about gold coloured human-shaped walking machines 1 . Instead
we are considering some-kind of platform or vehicle that moves through its environment
carrying some kind of payload. Almost without exception it is the payload that is of interest
- not the platform. However the vast majority of payloads require the host vehicle to navigate
—to be able to parameterize its position and surroundings and plan and execute trajectories
through it. Consider some typical autonomous vehicle and payloads:
• Humans in a airplane on autopilot (or car in the near-ish future. CMU NavLab project)
6
7
All of the above require navigation. This course will hopefully give you some insight into
how this can be achieved.
It is worth enumerating in general terms what makes autonomous navigation so hard. The
primary reason is that the majority of mobile robots are required to work in un-engineered
environments. Compare the work-spaces of a welding robot in automotive plant to one that
delivers blood samples between labs in a hospital. The former operates in a highly controlled,
known, time invariant (apart from the thing being built) scene. If computer vision is used
as a sensor then the workspace can be lit arbitrarily well to mitigate against shadows and
color ambiguity. Many industrial robots work in such well known engineered environments
that very little external sensing is needed — they can do their job simply by controlling
their own internal joint angles. Hospitals are a different ball-park altogether. The corridors
are dynamic — filling and emptying (eventually) with people on stretchers. Even if the
robot is endowed with a map of the hospital and fitted with an upward looking camera to
navigate off markers on the ceiling it still has to avoid fast moving obstacles (humans) while
moving purposely towards it’s goal destination. The more generic case involves coping with
substantial scene changes (accidental or malicious ) — for example doors closing in corridors
or furniture being moved. The thing then, that makes mobile robotics so challenging is
uncertainty. Uncertainty is pervasive in this area and we must embrace it to make progress....
7
Topic 2
A basic requirement of a mobile autonomous vehicle is path planning. With the vehicle in an
arbitrary initial position A we wish to issue a desired goal position B (including orientation)
and have the vehicle execute a trajectory to reach B . This sounds pretty simple and we can
think of several ways in which we could combine simple control laws that will get us from A
to B 1 Unfortunately the waters quickly become muddied when we start talking about our
other concerns:
• while executing its trajectory the vehicle must not smash into objects in the environ-
ment (especially true regarding squishy humans).
• we cannot guarantee that the vehicle in question can turn-on-the-spot and would like
to be able to operate a vehicle of arbitrary shape. These are called “kinematic” con-
straints.
• we expect only uncertain estimates of the location of the robot and objects in the
environment.
8
Holonomicity 9
Figure 2.1: Two holonomic vehicles. The underwater vehicle (ODIN, University of Hawaii)
can move in any direction irrespective of pose and the complex wheel robot PPRK (CMU)
driven by a “Palm Pilot” uses wheels that allow slip parallel to their axis of rotation. A sum
of translation and slip combine to achieve any motion irrespective of pose.
2.1 Holonomicity
Holonomicity is the term used to describe the locomotive properties of a vehicle with respect
to its workspace. We will introduce a mathematical definition of the term shortly but we
will begin by stating, in words, a definition:
• a car is non-holonomic : the global degrees of freedom are motion in x,y and heading
however locally, a car can only move forward or turn. It cannot slide sideways.(Even
the turning is coupled to motion).
9
Holonomicity 10
Figure 2.2: A commercial non-holonomic robot vehicle (Roombavac from iRobot coorpora-
tion. This vehicle can be purchased for about 100 USD - but is it’s utility great enough to
warrant the prices? Discuss....
• The “spherical” underwater robot (Odin) and the rolling wheel vehicle in Figure 2.1
are holonomic they can turn on the spot and translate instantaneously in any direction
without having to rotate first.
• A train is holonomic: it can move forwards or backwards along the track which is
parameterised by a single global degree of freedom — the distance along the track.
• The robot “Roombavac” (iRobot corporation) vacuum cleaner in Figure 2.1 is also
non-holonomic. It can rotate in place but cannot slide in any direction — it needs to
use a turn-translate-turn or turn-while-drive (like a car) paradigm to move.
It should be obvious to you that motion control for a holonomic vehicle is much easier
than for a non-holonomic vehicle. If this isn’t obvious consider the relative complexity of
parking a car in a tight space compared to driving a vehicle that can simply slide into the
space sideways (a hovercraft).
Unfortunately for us automation engineers, the vast majority of vehicles in use today
(i.e. used by humans) are non-holonomic. In fact intrinsically holonomic vehicles are so rare
and complex (or so simple 2 ) that we shall not discuss them further.
We can now place some formalism on our notion of holonomicity. We say a vehicle whose
state is parameterised by a vector x is non-holonomic if there exists a constraint Q such that
Q(x, ẋ, ẍ, · · · ) = 0 (2.1)
where the state derivatives cannot be integrated out. To illustrate such a constraint we will
take the case of a front wheel steered vehicle as shown in figure 2.1
2
path planning for a train is quite uninteresting
10
Configuration Space 11
x'
x =[x,y, θ] T
Immediately we can write a constraint expressing that the vehicle cannot slip sideways
that is a function of x and its first derivative:
ẋ − sin θ
ẏ . cos θ = 0 (2.2)
θ̇ 0
Q(x, ẋ) = 0 (2.3)
The state and its derivatives are inseparable and by our definition the vehicle is non-
holonomic.
We are describing the robot by its state x = [x1 , x2 , · · · xn ]T ( for a 2D plane vehicle commonly
x1 = x x2 = y x3 = θ )which is a n-state parameterisation. We call the space within which
x resides the configuration space C − space of the vehicle:
[
C= x (2.4)
∀x1 ,x2 ···xn
The configuration space ( or C − space ) is the set of all allowable configurations of the robot.
For a simple vehicle moving on a plane the configuration space has the same dimension as
the work space but for more complicated robots the dimension of the configuration space is
much higher. Consider the case of the bomb disposal vehicle in figure 2.2. The configuration
space for such a vehicle would be 11 — 3 for the base and another 8 for the pose of the arm
and gripper. We can view obstacles as defining regions of C − space that are forbidden. We
11
Configuration Space 12
Figure 2.4: A vehicle with a high dimensional configuration space - a commercial bomb-
disposal platform (picture courtesy of Roboprobe Ltd. The configuration space for a human
is immensely high dimensional —around 230.
can label this space as C⊗ and the remaining accessible/permissable space as C¯ such that
C = C⊗ ∪ C¯ . It is often possible to define and describe the boundaries of C⊗ (and hence the
boundaries of C¯ in which the vehicle must move) as a constraint equation. For example if
the workspace of a point robot in 2D x = [xy]T is bounded by a wall ax + by + c = 0 then
[
C¯ = {x | ax + by + c > 0} (2.5)
| {z }
union of all x for which ax+by+c > 0
Equation 2.6 simple states what configurations are open to the vehicle —nothing more. Any
path planning algorithm must guide the vehicle along a trajectory within this space while
satisfying any non-holonomic vehicle constraints and finally deliver the vehicle to the goal
pose. This clearly is a hard problem. Two poses that may be adjacent to each other in state
space may require an arbitrarily long path to be executed to transition between them. Take,
for example, the seemingly simple task of turning a vehicle through 180o when it is near
a wall. One solution trajectory to the problem is shown in figure 2.2. The non-holonomic
vehicle constraints conspire with the holonomic constraint imposed by the wall to require a
complicated solution trajectory.
12
The Minkowski-Sum 13
3 1 4
3
Figure 2.5: A simple task - turning through 180o quickly becomes complicated by
C − spaceconstraints.
Real robots have arbitrary shapes and these shapes make for complicated interactions with
obstacles which we would like to simplify. One way to do this is to transform the problem
to one in which the robot can be considered as a point-object and a technique called the
“Minkowski-Sum” does just this. The basic idea is to artificially inflate the extent of obstacles
to accommodate the worst-case pose of the robot in close proximity. This is easiest to
understand with a diagram shown in Figure 2.3. The idea is to replace each object with a
virtual object that is the union of all poses of the vehicle that touch the obstacle. Figure 2.3
has taken a conservative approach and “replaced” a triangular vehicle with a surrounding
circle. The minimal Minkowski-Sum would be the union of the obstacle and all vehicle poses
with the vehicle nose just touching it boundary. With the obstacles suitably inflated the
vehicle can be thought of a point-object and we have a guarantee that as long as it keeps to
the new, shrunken, free space it cannot hit an object 3 . Note it is usual to fit a polygonal
hull around the results of the Minkowski-Sum calculation to make ensuing path planning
calculations easier.
So now we have a method by which we can calculate C¯ . The next big question is how
exactly do we plan a path through it? How do we get from an initial pose to a goal pose?
We will consider three methods : Voronoi, “Bug” and Potential methods.
3
This doesn’t mean that awkward things won’t happen — the situation shown in figure 2.2 is still
possible. However progress can be made by planning motion such that at object boundaries the vehicle is
always capable of moving tangentially to the boundaries
13
Voronoi Methods 14
Figure 2.6: The Minkowski-Sum transforms a arbitrarily shaped vehicle to a point while
inflating the obstacle. The result is guaranteed free space outside the inflated object bound-
ary.
100
80
60
40
20
−20
−20 0 20 40 60 80 100 120 140 160
Figure 2.7: A Voronoi diagram for obstacle avoidance in the presence of point objects.
Voronoi diagrams are elegant geometric constructions 4 that find applications throughout
computer science — one is shown in figure 2.4. Points on a 2D-Voronoi diagram are equi-
distant from the nearest two objects in the real world. So the Voronoi diagram of two points
a, b is the line bisecting them — all points on that line are equidistant from a, b. The efficient
computation of Voronoi diagrams can be quite a complex matter but what is important
here is that algorithms exist that when given a set of polygonal objects can generate the
appropriate equi-distant loci. So how does this help us with path planning? Well, if we follow
the paths defined by a Voronoi diagram we are guaranteed to stay maximally far away from
4
http://www.voronoi.com/
14
Bug Methods 15
nearby objects. We can search the set of points on the diagram to find points that are closest
to and visible from the start and goal positions. We initially drive to the “highway entry”
point, follow the “Voronoi - highways” until we reach the “exit point” where we leave the
highway and drive directly towards the goal point.
The generation of a global Voronoi diagram requires upfront knowledge of the environment.
In many cases this is unrealistic. Also Voronoi planners by definition keep the vehicle as far
away from objects as possible - this can have two side effects. Firstly the robot may be using
the objects to localise and moving away from them makes them harder to sense. Secondly
the paths generated from a Voronoi planner can be extremely long and far from the shortest
path (try playing with the Matlab Voronoi function). An alternative family of approaches
go under the name of “ bug algorithms”. The basic algorithm is simple:
1. starting from A and given the coordinates of a goal pose B draw a line AB (it may
pass through obstacles that are known or as yet unknown)
2. move along this line until either the goal is reached or an obstacle is hit.
3. on hitting an obstacle circumnavigate its perimeter until AB is met
4. goto 2
In contrast to the Voronoi approach this method keeps the vehicle as close to the obstacles
as possible (but we won’t hit them as the obstacles have been modified by the Minkowski
sum!). However the path length could be stupidly long. A smarter modification would be
to replace step 3 in the original algorithm with “on hitting and obstacle circumnavigate it’s
perimeter until AB is met or the line AB becomes visible in which case head for a point on
AB closer to B ” Figure 2.5 shows the kind of trajectory this modified “VisBug” algorithm
would execute. Clearly the hugging the object boundary is not always a good plan but
interestingly it is guaranteed to get the robot to the goal location if it is indeed reachable.
A third very popular method of path planning is a so called “Potential Method”. Once again
the idea is very simple. We view the goal pose as a point of low potential energy and the
15
Potential Methods 16
goal
start
Figure 2.8: The visbug algorithm executing a goal-seek trajectory around Southern Italy.
obstacles as areas of high potential energy. If we think of the vehicle as a ball-bearing free
to roll around on a “potential terrain“ then it will naturally roll around obstacles and down
towards the goal point. The local curvature and magnitude of the potential field can be used
to deduce a locally preferred motion. We now firm up this intuitive description with some
mathematics.
At any point x we can write the total potential UΣ as a sum of the potential induced Uo
by k obstacles and the potential induced by the goal Ug :
X
UΣ (x) = Uo,i (x) + Ug (x) (2.7)
i=1:k
Now we know that the force F(x) exerted on a particle in a potential field U Σ (x) can be
written as :
16
Potential Methods 17
The next question is what exactly do the potential functions look like. Well, there is
no single answer to that — you can “roll your own”. A good choice would be to make the
potential of an obstacle be an inverse square function of the distance between vehicle and
obstacle. We may also choose an everywhere-convex potential for the goal so that where ever
we start, in the absence of obstacles, we will “fall” towards the goal point. Figure 2.6 shows
two useful potential candidates (forgive the pun). Defining ρ(x) as the shortest distance
goal potential
Figure 2.9: Two typical potential functions - inverse quadratic for obstacle and quadratic
for the goal.
between the obstacle and the vehicle (at x ) and xg as the goal point, the algebraic functions
for these potentials are:
³ ´2
1 1 1
− ∀ ρ(x) ≤ ρ0
Uo,i (x) = η 2 ρ(x) ρ0 (2.10)
0 otherwise
1
Ug (x) = (x − xg )2 (2.11)
2
The term ρ0 places a limit on the region of space affected by the potential field — the virtual
vehicle is only affected when it comes with ρ0 of an obstacle. η is just a scale factor. Simple
differentiation allows us to write the force vector exerted on a virtual point vehicle as:
( ³ ´
1 1 1 ∂ρ(x)
η ρ(x) − ρ0 ρ(x) 2 ∂x ∀ρ(x) ≤ ρ0
Fo,i = (2.12)
0 otherwise
where ∂ρ(x)
∂x
is the vector of derivatives of the distance function ρ(x) with respect to x, y, z.
We proceed similarly for the goal potential. So to figure out the force acting on a virtual
vehicle and hence the direction the real vehicle should move in, we take the sum of all obstacle
17
Potential Methods 18
forces and the goal force – a very simple algorithm. The course web-site has example code
for a potential path planner written in Matlab. Download it and play.
Although elegant, the potential field method has one serious drawback — it only acts
locally. There is no global path planning at play here — the vehicle simply reacts to local
obstacles, always moving in a direction of decreasing potential. This policy can lead to the
vehicle getting trapped in a local minimum as shown in figure 2.6. Here the vehicle will
descend into a local minima in front of the two features and will stop. Any other motion will
increase its potential. If you run the code you will see the vehicle getting stuck between two
features from time to time 6 . The potential method is someway between the bug method
and Voronoi method in terms of distance to obstacles. The bug algorithm sticks close to
them, the Voronoi as far away as possible. The potential method simply directs the vehicle
in a straight line to a goal unless it moves into the vicinity of of an obstacle in which case it
is deflected.
500
GOAL
400
300
200 TRAP
3
100
2.5
2 2
2
1.5
1 1
1.5
0.5
0 TRAP GOAL
0
1
−1 −0.5
−1 0.5
−2 −1.5
0 0.5 1 1.5 2 2.5 3
−2 0
Figure 2.10: A pathological potential terrain and goal combination. The vehicle can get
stuck in a local-minima or “trap”
6
one way out of this is to detect a trap and temporarily move the goal point, the problem is, to where?
18
Topic 3
3.1 Introduction
This lecture will begin to cover a topic central to mobile robotics — Estimation. There
is a vast literature on Estimation Theory encompassing a rich variation of techniques and
ideas. We shall focus on some of the most commonly used techniques which will provide the
tools to undertake a surprisingly diverse set of problems. We shall bring them to bear on
real life mobile-robot problems. Although we shall discuss and illustrate these techniques
in a domain specific fashion, you will do well to keep in your mind that these are general
techniques and can be applied to loads of other non-robotic problems1 .
To make sure we are all on the same page it is worth stating (in a wordy fashion) what we
mean by “Estimation”:
19
What is Estimation? 20
There is nothing surprising there —it fits intuitively with our everyday meaning of the
word. However, we may have our interest piqued when we impose the following additional
characteristics to an estimator.
• We expect the measurements to be noise corrupted and would expect to see this un-
certainty in input transformed to uncertainty in inference.
• We do not expect the measurements to always be of x directly. For example a GPS
receiver processes measurements of time (4 or more satellites transmit “time-of-day-
at-source”) and yet it infers Euclidean position.
• We might like to incorporate prior information in our estimation. For example we may
know the depth and altitude of a submarine but not its latitude.
• We might like to employ a model that tells us how we expect a system to evolve over
time. For example given a speed and heading we could reasonably offer a model on
how a ship might move between two time epochs — open-loop.
• We expect these models to be uncertain. Reasonably we expect this modelling error
to be handled in the estimation process and manifest itself in the uncertainty of the
estimated state.
Pretty quickly we see that if we can build an estimator capable of handling the above
conditions and requirements we are equipping ourselves with a powerful and immensely
useful inference mechanism.
Uncertainty is central to the problem of estimation (and robotics). After all, if it weren’t
for uncertainty many problems would have a simple algebraic solution —“give me the dis-
tances to three known points and the algebraically-determined intersection of three circles
will define my location”. We will be trying to find answers to more realistic questions like
“I have three measurements (all with different uncertainty) to three surveyed points (known
roughly) — what is your best estimate of my location?”. Clearly this is a much harder
and more sophisticated question. But equipped with basic probability theory and a little
calculus the next sections will derive techniques capable of answering these questions using
a few modest assumptions.
Probability theory and random-variables are the obvious way to mathematically manipulate
and model uncertainties and as such much of this lecture will rely on your basic knowledge
20
Maximum Likelihood Estimation 21
We want to obtain our best estimate x̂ for a parameter x given a set of k measurements
k
Z = {z1 , z2 · · · zk }. We will use a “hat” to denote an estimated quantity and allow the
absence of a hat to indicate the true (and unknowable) state of the variable.
It is sane to suppose that a measurement z that we are given is in some way related to
the state we wish to estimate 2 . We also suppose that measurements (also referred to as
observations in these notes) are not precise and are noise corrupted. We can encapsulate
both the relational and uncertain aspects by defining a likelihood function:
L , p(z|x) (3.1)
The distribution p(z|x) is the conditional probability of the measurement z given a particular
value of x. Figure 3.3 is just such a distribution — a Gaussian in this case( C is just a
normalising constant):
1 1 T −1
p(z|x) = e− 2 (z−x) P (z−x) (3.2)
C
Notice that equation 3.2 is a function of both x and z . Crucially we interpret L as func-
tion of x and not z as you might initially think. Imagine we have been given an observation
z and an associated pdf L for which we have a functional form of (Equation 3.2). We form
our maximum likelihood estimate x̂m.l by varying x until we find the maximum likelihood.
clearly, as if they were independent Zk would contain no information about x and the sensor providing
2
21
Maximum A-Posteriori - Estimation 22
0.1
0.09
0.08
0.07 z=15.000000
0.06
0.05
Λ
0.04
0.03
0.02
0.01
0
0 5 10 15 20 25 30 35
x
In many cases we will already have some prior knowledge on x . Imagine x is a random
variable which we have good reason to suppose is distributed as p(x ). For example, perhaps
we know the intended (by design) rotational velocity µp of a CDROM drive - we might model
our prior knowledge of this as a gaussian ∼ N (µp , σp2 ). If we have a sensor that can produce
an observation z of actual rotation speed with a behaviour modelled as p(z|x) we can use
Bayes rule to come up with a posterior pdf p(x|z) that incorporates not only the information
from the sensor but also our assumed prior knowledge on x :
p(z|x)p(x)
p(x|z) =
p(z)
= C × p(z|x)p(x)
The maximum a-posteriori - MAP finds the value of x which maximises p(z|x)p(x) (the
normalising constant is not a function of x ).
22
Maximum A-Posteriori - Estimation 23
Lets quickly finish the CD example, assume our sensor also has a gaussian likelihood
function, centered on the observation z but with a variance σz2 :
(x − µp )2
p(x) = C1 exp{− }
2σp2
(z − x)2
p(z|x) = C2 exp{− }
2σz2
p(z|x)p(x)
p(z|x) =
p(z)
= C(z) × p(z|x) × p(x)
(x − µp )2 (z − x)2
= C(z) exp{− − }
2σp2 2σz2
p(z|x) is maximum when the exponent is zero. A simple way to find what value of x achieves
this is to re-write the exponent of p(z|x) as
(x − α)2 (x − µp )2 (z − x)2
= − − (3.5)
β2 2σp2 2σz2
Expanding the R.H.S and comparing coefficients swiftly leads to:
σz2 µp + σp2 z
α=
σz2 + σp2
σz2 σp2
β2 =
σz2 + σp2
σ 2 µ +σ 2 z σ2 σ2
Therefore the MAP estimate x̂ is zσ2p+σ2p and it has variance σ2z+σp2 . It is interesting and
z p z p
informative to note that as we increase the prior uncertainty in the CD speed by increasing
σp towards infinity then x̂ tends towards z —the ML estimate. In other words when we have
an uninformative prior the MAP estimate is the same as the ML estimate. This makes sense
as in this case the only thing providing information is the sensor (via its likelihood function).
There several other things that you should be sure you appreciate before progressing:
2
Decreasing posterior variance Calling the posterior variance σ+ ( which is β 2 above),
we note it is smaller than that of the prior. In the simple example above this can
be seen as σ12 = β12 = σ12 + σ12 . This is not surprising as the measurement is adding
+ p z
3
information
3
Indeed a useful metric of information is closely related to the inverse of variance — note the addition of
inverses...
23
Minimum Mean Squared Error Estimation 24
Update to prior In the Gaussian example above we can also write x̂map as an adjustment
to the prior mean:
σp2
x̂map = µp + 2 × (z − µp ) (3.6)
σp + σz2
clearly if the observation matches the mode of the prior (expected) distribution then
x̂map is unchanged from the prior (but it’s uncertainty still decreases). Also note that
the correction z − µp is scaled by the relative uncertainty in both prior and observation
- if the observation variance σz2 is huge compared to σp2 (i.e. the sensor is pretty
terrible) then irrespective of the magnitude of z − µp , x̂map is still µp and the decrease
in uncertainty is small. This too makes sense — if a sensor is very noisy we should
not pay too much attention to disparities between expected (the prior) and measured
values. This is a concept that we will meet again soon when we discuss Kalman
Filtering.
Another key technique for estimating the value of a random variable x is that of minimum
mean squared error estimation. Here we assume we have been furnished with a set of
observations Zk . We define the following cost function which we try to minimise as a
function of x̂ :
x̂mmse = arg min E{(x̂ − x)T (x̂ − x)|Zk } (3.7)
x̂
The motivation is clear, we want to find an estimate of x that given all the measurement
minimises the expected value of sum of the squared errors between the truth and estimate.
Note also that we are trying to minimise a scalar quantity.
24
Recursive Bayesian Estimation 25
Why is this important? - it tells us that the mmse estimate of a random variable
given a whole bunch of measurements is just the mean of that variable conditioned on the
measurements. We shall use this result time and time again in coming derivations.
Why is it different from the Least Squares Estimator? They are related (the
LSQ estimator can be derived from Bayes rule) but here x is a random variable where in
the LSQ case x was a constant unknown. Note we haven’t discussed the LSQ estimator yet
- but by the time you come to revise from these notes you will have.
The idea of MAP estimation leads naturally to that of recursive Bayesian estimation. In
MAP estimation we fused both a-priori beliefs and current measurements to come up with
an estimate, x̂ , of the underlying world state x . If we then took another measurement
we could use our previous x̂ as the prior, incorporate the new measurement and come up
with a fresh posterior based now, on two observations. The appeal of this approach to a
robotics application is obvious. A sensor (laser radar etc) produces a time-stamped sequence
of observations. At each time step k we would like to obtain an estimate for it’s state given
all observations up to that time ( the set Zk ). We shall now use Bayes rule to frame this
more precisely:
and also
25
Recursive Bayesian Estimation 26
so
if we assume (reasonably) that given the underlying state, observations are conditionally
independent we can write
p(x|Zk−1 )p(Zk−1 )
p(Zk−1 |x) = (3.20)
p(x)
p(x|Zk−1 )p(Zk−1 )
p(x|Zk )p(Zk ) = p(zk |x) p(x) (3.21)
p(x)
= p(zk |x)p(x|Zk−1 )p(Zk−1 ) (3.22)
so
note that
so
p(Zk−1 ) 1
= (3.26)
p(Zk ) p(zk |Zk−1 )
26
Recursive Bayesian Estimation 27
p(zk |x)p(x|Zk−1 )
p(x|Zk ) =
p(zk |Zk−1 )
(3.27)
So what does this tell us? Well, we recognise p(x|Zk ) as our goal - the pdf of x conditioned
on all observations we have received up to and including time k. The p(zk |x) term is just the
likelihood of the k th measurement. Finally p(x|Zk−1 ) is a prior — it is our last best estimate
of x at time k − 1 which at that time was conditioned on all the k − 1 measurements that had
been made up until that time. The recursive Bayesian estimator is a powerful mechanism
allowing new information to be added simply by multiplying a prior by a (current) likelihood.
Note that nothing special has been said about the form of the distributions manipulated
in the above derivation. The relationships are true whatever the form of the pdfs. However
we can arrive at a very useful result if we consider the case where we assume Gaussian priors
and likelihoods... the linear Kalman Filter.
27
Topic 4
4.1 Motivation
z = Hx (4.1)
We wish to take this data and solve this equation to find x in terms of z . Initially you
may naively think that a valid solution is
x = H−1 z (4.2)
which is only a valid solution if H is a square matrix with | H |6= 0 — H must be invertible.
We can get around this problem by seeking a solution x̂ that is closest to the ideal1 The
metric of “closest” we choose is the following:
x̂ = arg min || Hx − z ||2 (4.3)
x
© ª
x̂ = arg min (Hx − z)T (Hx − z) (4.4)
x
Equation 4.4 can be seen to be a “least squares” criterion. There are several ways to solve
this problem we will describe two of them - one appealing to geometry and one to a little
calculus.
1
For this section we will assume that we have more observations than required ie dim(z) > dim(x) which
assures that there is a unique “best” solution
28
Motivation 29
Recall from basic linear algebra that the vector Hx is a linear sum of the columns of H. In
other words Hx ranges over the column space of H. We seek a vector x such that Hx is
closest to the data vector z . This is achieved when the error vector e = Hx−z is orthogonal
to the space in which Hx is embedded. Thus e must be orthogonal to every column of H:
HT (Hx − z) = 0 (4.5)
HT Hx = HT z (4.6)
T −1 T
x = (H H) H z (4.7)
This is the least squares solution for x . The matrix (HT H)−1 HT is called the pseudo-inverse
of H.
The least squares solution for the linear system Ax = b with Am,n m > n is
29
Weighted Least Squares 30
Imagine now that we have some information regarding how reliable each of the elements in
z is. We might express this information as a diagonal measurement covariance matrix R :
2
σz1 0 0
2
R = 0 σz2 · · ·
(4.12)
.. .. . .
. . .
It would be natural to weight each element of the error vector e according to our uncertainty
in each element of the measurement vector z - ie by R−1 . The new minimisation becomes:
x̂ = arg min || R−1 (Hx − z) ||2 (4.13)
x
Carrying this through the same analysis yields the weighted linear least squares esti-
mate:
x̂ = (HT R−1 H)−1 HR−1 z (4.14)
The previous section allows us to derive a least squares estimate under a linear observation
model. However most interesting problems will involve non-linear models - measuring the
Euclidean distance between two points for example. We now have a new minimisation task:
x̂ = arg min || h(x) − z ||2 (4.15)
x
where
∂h ∂h1
1
∂x1
··· ∂xm
∂h . ..
∇Hx0 = = .. . (4.19)
∂x ∂hn ∂hn
∂x1
··· ∂xm
| {z }
evaluated at x0
30
Weighted Least Squares 31
Equation 4.18 can be seen to be a linear least square problem - something we have already
solved and stated in equation 4.11. By inspection then we can write an expression for δ that
minimises the right hand side of 4.18:
¡ ¢−1
δ = ∇Hx0 T ∇Hx0 ∇Hx0 T [z − h(x0 )] (4.20)
We now set x1 = x0 +δ and iterate again until the norm falls below a tolerance value. Like the
linear case, there is a natural extension to the weighted non linear case. For a measurement
z with variance R the weighted non-linear least squares algorithm is as follows:
2. Evaluate ¡ ¢−1
δ = ∇Hx̂ T R−1 ∇Hx̂ ∇Hx̂ T R−1 [z − h(x̂)]
3. Set x̂ = x̂ + δ
So why is the non-linear LSQ problem interesting? Well, non-linear least squares
allows us to solve some interesting and realistic navigation problems. For example, consider
the case of an autonomous underwater vehicle (AUV) moving within a network of acoustic
beacons. The AUV shown in figure 4.2.2 is about to be launched on a mine detection mission
in the Mediterranean. Within the hull (which floods) is a transceiver which emits a “call”
pulse into the water column. Beacons deployed at known locations detect this pulse and
reply with “acknowledge” pulses which are detected by the transceiver. The difference in
time between “call” and “acknowledge” is proportional to the distance between vehicle and
each beacon. Figure 4.2.2 shows a diagram of this kind of navigation (which is very similar
to how GPS works).
31
Weighted Least Squares 32
Figure 4.1: A Long Baseline Acoustic Network for an Autonomous Underwater Vehicle
(AUV)
so we can stack them into a single vector2 . The model of the long-baseline transceiver
operating in water with speed of sound c can be written as follows:
£ ¤T
z = t1 t2 t3 t4 = h(xv ) (4.21)
|| xb1 − xv ||
t1
t2 = 2 || xb2 − xv ||
(4.22)
c || xb3 − xv ||
t3
|| xb4 − xv ||
so
∆x1 ∆y1 ∆y1
2 ∆ x2 ∆y2 ∆y2
∇Hxv =− (4.23)
rc ∆x3 ∆y3 ∆y3
∆x4 ∆y4 ∆y4
where
∆xi = xbi − x
∆yi = ybi − y
∆zi = zbi − z
p
r = (xbi − x)2 ) + (ybi − y)2 ) + (zbi − z)2 )
2
In practice of course the signal that travels the furthest comes in last - the measurements are staggered.
This is only a problem if the vehicle is moving
32
Weighted Least Squares 33
Figure 4.2: An autonomous underwater vehicle about to be launched. The large fork on the
front is a sonar designed to detect buried mines. Within the hull is a transceiver which emits
a “call” pulse into the water column. Beacons deployed at known locations detect this pulse
and reply with a “acknowledge” pulses which are detected by the transceiver. The difference
in time between “call” and “acknowledge” is proportional to the distance between vehicle
and each beacon. (photo courtesy of MIT) see figure 4.2.2
With the jacobian calculated and given a set of four measurements to the beacons we
can iteratively apply the non-linear least squares algorithm until the change in x v between
iterations becomes small enough to disregard. You might be surprised that many of the
oil rigs floating in the Gulf of Mexico basically use this method to calculate their position
relative to the well-head thousands of meters below them. Oil rigs are perhaps one of the
largest mobile-robots you are likely to encounter.
One issue with the Least Squares solution is that enough data has to be accumulated to
make the system observable - HT H must be invertible. For example, in our subsea example,
acoustic-refractive properties of the water column may mean that replies from a particular
beacon are never detected. Alternatively acoustic noise may obliterate the true signal leading
to false detections. Figure 4.2.2 shows some real LBL data from vehicle shown in Figure
4.2.2. Most of the noise is due to the enormously powerful fork-like sonar on its nose.
The Kalman filter which we shall discuss, derive and use shortly is one way to overcome
this problem of instantaneous un-observability.
33
Weighted Least Squares 34
Raw Data
3000
TOF x C (s)
2000
1000
0
0 500 1000 1500 2000 2500 3000
Data Sketch
1500
SAS turns on
TOF x C (m)
1000
500
0
0 500 1000 1500 2000 2500 3000
Recovered Data
1500
TOF x C (m)
1000
500
0
0 500 1000 1500 2000 2500 3000
Time (s)
Figure 4.3: Some real data from the ocean. The synthetic aperture adds a lot of noise to the
LBL sensing scheme. However it is possible to separate true signal from noise as shown in
the lower graphs. The y axes are time of flight × c giving effective distance between vehicle
and beacons.
34
Topic 5
Keep in mind
The following algebra is simple but a little on the laborious side and was provided here
for completeness. The following section will involve a linear progression of simple algebraic
steps. I hope you will derive some degree of intellectual satisfaction in seeing the equations
being derived using only Bayes rule. Of course, like many things in engineering, derivations
need not be done at every invocation of a smart idea but it is important to understand the
underlying principles. The exams will not ask for a complete derivation of the Kalman filter
but may ask you to explain the underlying concepts with reference to the key equations.
We will now assume that the likelihood p(z|x) and a prior p(x) on x are Gaussian. Further-
more, initially, we are going to model our sensor as something that produces a observation
z which is a noise corrupted linear function of the state x :
z = Hx + w (5.1)
Here w is a gaussian noise with zero mean and covariance R so that
1 1
p(w) = n/2 1/2
exp{− wT R−1 w}. (5.2)
(2π) | R | 2
35
The Linear Kalman Filter 36
Note that up until now the examples have often been dealing with scalar 1D distributions.
Here we generalise to the multidimensional case but things should still look familiar to you.
We shall let the state have nx dimensions and the observation vector have nz dimensions.
We can also use a multidimensional Gaussian with mean xª and covariance Pª to describe
our prior belief in x :
1 1
p(x) = exp{− (x − xª )T P−1
ª (x − xª )} (5.4)
(2π)nx/2 | Pª | 1/2 2
.
Now we can use Bayes rule to figure out an expression for the posterior p(x|z).:
p(z|x)p(x)
p(x|z) = (5.5)
p(z)
p(z|x)p(x)
= R∞ (5.6)
−∞
p(z|x)p(x)dx
1
(2π)nz/2 |R|1/2
exp{− 12 (z − Hx)T R−1 (z − Hx)} (2π)nx/21|Pª |1/2 exp{− 12 (x − xª )T P−1
ª (x − xª )}
=
C(z)
(5.7)
This looks pretty formidable but we do know that we will end up with a gaussian (gaussian
× gaussian = gaussian) and scale factors are not important therefore we can disregard the
denominator and focus on the product:
1 1
exp{− (z − Hx)T R−1 (z − Hx)} exp{− (x − xª )T P−1
ª (x − xª )} (5.8)
2 2
or equivalently:
¡ ¢
exp{−1/2 (z − Hx)T R−1 (z − Hx) + (x − xª )P−1
ª (x − xª )
T
} (5.9)
We know (because gaussian × gaussian = gaussian) that we can find a way to express the
above exponent in a quadratic way:
(x − x⊕ )T P−1
⊕ (x − x⊕ ) (5.10)
36
The Linear Kalman Filter 37
We can figure out the new mean x⊕ and covariance P⊕ by expanding expression 5.9 and
comparing terms with expression 5.10. Remember we want to find the new mean because
Equation 3.14 tells us this will be the MMSE estimate. So expanding 5.9 we have:
xT P−1 T −1 T −1 T −1 T T −1 T T −1 −1 T −1
ª x−xª Pª xª −x Pª xª +xª Pª xª +x H R Hx−x H R z−zR Hx+z R z
(5.11)
(5.12)
Expanding 5.10:
xT P−1 T −1 T −1 T −1
⊕ x − x P⊕ x ⊕ − x ⊕ P⊕ x + x ⊕ P⊕ x ⊕ . (5.13)
Comparing first terms in 5.12 and 5.13 we immediately see that
P⊕ = (P−1 T −1 −1
ª + H R H) . (5.14)
Comparing the second terms we see that:
P−1 −1 T −1
⊕ x⊕ = Pª xª + H R z. (5.15)
Therefore we can write the MMSE estimate, x⊕ as
x⊕ = (P−1 T −1 −1 −1 T −1
ª + H R H) (Pª xª + H R z). (5.16)
We can combine this result with our understanding of the recursive Bayesian filter we
covered in section 3.6. Every time a new measurement becomes available we update our
estimate and its covariance using the above two equations.
There is something about the above two equations 5.14 and 5.16 that may make them
inconvenient — we have to keep inverting our prior covariance matrix which may be com-
putationally expensive if the state-space is large 1 . Fortunately we can do some algebra to
come up with equivalent equations that do not involve an inverse.
We begin by stating a block matrix identity. Given matrices A , B and C the following
is true (for non-singular A ):
(A + BCBT )−1 = A−1 − A−1 B(C−1 + BT A−1 B)−1 BT A−1 (5.17)
1
in some navigation applications the dimension of x can approach the high hundreds
37
The Linear Kalman Filter 38
or
= (I − WH)Pª (5.21)
where
S = HPª HT + R (5.22)
T −1
W = Pª H S (5.23)
Now look at the form of the update equation 5.16 it is a linear combination of x and z .
Combining 5.20 with 5.16 we have:
x⊕ = (P−1 T −1 −1 −1 T −1
ª + H R H) (Pª xª + H R z) (5.24)
= (P−1
ª
T −1 −1 T
+ H R H) (H R z) + (I − −1
WH)Pª (P−1
ª xª ) (5.25)
−1
= + HT R−1 H)−1 (HT R−1 z) + xª + W(−Hxª )
(Pª (5.26)
= Cz + xª + W(−Hxª ) (5.27)
where
C = (P−1 T −1 −1 T −1
ª + H R H) H R (5.28)
Taking a step aside we note that both
HT R−1 (HPª HT + R) = HT R−1 HPª HT + HT (5.29)
and also
(P−1 T −1 T T T −1
ª + H R H)Pª H = H + H R HPª H
T
(5.30)
so
(P−1 T −1 −1 T −1
ª + H R H) H R = Pª HT (HPª HT + R)−1 (5.31)
therefore
C = Pª HT S−1 (5.32)
= Wfrom 5.23 (5.33)
38
The Linear Kalman Filter 39
We can now summarise the update equations for the Linear Kalman Filter:
x⊕ = xª + Wν
P⊕ = Pª − WSWT
ν = z − Hxª
S = HPª HT + R
W = Pª HT S−1
The previous section showed how an observation vector (z ) related in some linear way to
a state we wish to estimate (x ) can be used to update in a optimal way a prior belief
/estimate. This analysis lead us to the so called Kalman update equations. However there
is another case which we would like to analyse: given a prior at time k − 1, a (imperfect)
model of a system that models the transition of state from time k − 1 to k, what is the new
estimate at time k?
39
The Linear Kalman Filter 40
Mathematically we are being told that the true state behaves in the following fashion:
Here we are modelling our uncertainty in our model by adding a linear transformation
of a “white noise” 2 term v with strength (covariance) Q. The term u(k) is the control
vector at time k. It models actions taken on the plant that are independent of the state
vector itself. For example, it may model an acceleration on a model that assumes constant
velocity. (A good example of u is the steer angle on a car input into the system by a driver
(machine or human))
Our general approach will be based on an inductive argument. However first we shall intro-
duce a little more notation. We have already shown (and used) the fact that given a set Z k
of k observations the MMSE estimate of x is
We now drop the “mmse” subscript and start to use two parenthesised time indexes i and
j. We use x̂(i|j) to mean the “MMSE estimate of x at time i given observations up until
time j”. Incorporating this into Equation 5.36 we have
2
a random variable distributed according to a zero-mean gaussian pdf
3
If this sounds confusing focus on the fact that we are maintaining a probability distribution for x . Our
estimate at time i using measurements up until time j (x̂(i|j)) is simply the mean of this distribution and
it has variance P(i|j). If our distribution is in fact a Gaussian then these are the only statistics we need to
fully characterize the pdf. This is exactly what a Kalman filter does — it maintains the statistics of a pdf
that “best” represents x given a set of measurements and control inputs
40
The Linear Kalman Filter 41
Now back to our question about incorporating a plant model F and a control signal u.
Imagine at time k we have been provided with MMSE estimate of x (k-1). Using our new
notation we write this estimate and its covariance as x̂(k − 1|k − 1) and P(k − 1|k − 1).
We are also provided with a control signal u(k) we want to know how to figure out a new
estimate x̂(k|k − 1) at time k. Note that we are still only using measurements up until time
(k − 1) ( i > j in Equation 5.37) however we are talking about an estimate at time k. By
convention we call this kind of estimate a “prediction”. You can think of predicting as an
“open loop” step as no measurements of state are used to correct the calculation.
This is a both simple and intuitive result. It simply says that the best estimate at time k
given measurements up until time k − 1 is simply the projection of the last best estimate
x̂(k − 1|k − 1) through the plant model. Now we turn our attention to the propagation of
the covariance matrix through the prediction step.
41
The Linear Kalman Filter 42
so
Now we are going to make a moderate assumption: that the previous estimate x̂(k − 1|k − 1)
and the noise vector (modelling inaccuracies in either model or control) are uncorrelated (i.e.
E{(x(k − 1) − x̂(k − 1|k − 1))v(k)T } is the zero matrix). This means that when we expand
Equation 5.47 all cross terms between x̂(k − 1|k − 1) and v(k) disappear. So:
P(k|k − 1) = FE{(x(k − 1) − x̂(k − 1|k − 1))(x(k − 1) − x̂(k − 1|k − 1))T |Zk−1 }FT +
(5.48)
GE{v(k)v(k)T |Zk−1 }GT (5.49)
T T
P(k|k − 1) = FP(k − 1|k − 1)F + GQG (5.50)
This result should also seem familiar to you ( remember that if x ∼ N (µ, Σ) and y = M x
then y ∼ N (M µ, M ΣM T ) ?).
We are now almost in a position to put all the pieces together. To start with we insert our
now temporal-conditional notation into the Kalman update equations. We use x̂(k|k − 1) as
the prior xª and process an observation z at time k. The posterior x⊕ becomes x̂(k|k)
So now we are in a position to write down the “standard Linear Kalman Filter Equations”.
If the previous pages of maths have started to haze your concentration wake up now as you
will need to know and appreciate the following:
42
The Linear Kalman Filter 43
prediction:
update:
where
5.1.3 Discussion
There are several characteristics of the Kalman Filter which you should be familiar with
and understand well. A firm grasp of these will make your task of implementing a KF for a
robotics problem much easier.
Recursion The Kalman Filter is recursive. The output of one iteration becomes the input
to the next.
Initialising Initially you will have to provide P(0|0) and x̂(0|0). These are initial guesses
(hopefully derived with some good judgement)
Structure The Kalman filter has a predictor-corrector architecture. The prediction step
is corrected by fusion of a measurement. Note that the innovation ν is a difference
between the actual observation and the predicted observation (Hx̂(k|k − 1)). If these
43
The Linear Kalman Filter 44
two are identical then there is no change to the prediction in the update step —both
observation and prediction are in perfect agreement. Fusion happens in data space.
Asynchronisity The update step need not happen at every iteration of the filter. If at a
given time step no observations are available then the best estimate at time k is simply
the prediction x̂(k|k − 1).
Prediction Covariance Inflation Each predict step inflates the covariance matrix — you
can see this by the addition of GQGT . This makes sense sice we are only using a
pre-conceived idea of how x will evolve. By our own admission — the presence of a
noise vector in the model — the model is inaccurate and so we would expect certainty
to dilate (uncertainty increase) with each prediction.
Update Covariance Deflation Each update step generally4 deflates the covariance ma-
trix. The subtraction of WSWT 5 during each update illustrates this. This too makes
sense. Each observation fused contains some information and this is “added” to the
state estimate at each update. A metric of information is related to the inverse of
covariance — note how in equation 5.16 HT R−1 H is added to the inverse of P, this
might suggest to you that information is additive. Indeed, equation 5.16 is the so called
“information form” of the Kalman Filter.
Observability The measurement z need not fully determine the state x (ie in general H
is not invertible). This is crucial and useful. In a least squares problem at every
update there have to be enough measurements to solve for the state x , However the
Kalman filter can perform updates with only partial measurements. However to get
useful results over time the system needs to be observable otherwise the uncertainty in
unobserved state components will grow without bound. Two factors make this possible
: the fact that the priors presumably contain some information about the unobserved
states (they were observed in some previous epoch) and the role of correlations.
Correlations The covariance matrix P is highly informative. The diagonals are the princi-
pal uncertainties in each of the state vector elements. The off diagonals tell us about
the relationships between elements of our estimated vector x̂ — how they are corre-
lated. The Kalman filter implicitly uses these correlations to update states that are not
observed directly by the measurement model! Let’s take a real life example. Imagine
we have a model [F, B Q] of a plane flying. The model will explain how the plane
moves between epochs as a function of both state and control. For example at 100m/s,
nose down 10 deg, after 100ms the plane will have travelled 10m forward (y direction)
and perhaps 1.5 m down (in z direction). Clearly the changes and uncertainties in y
and z are correlated — we do not expect massive changes in height for little change
4
technically it never increases it
5
which will always be positive semi definite
44
The Linear Kalman Filter 45
Prediction
no available ? yes
S = H P(k|k-1) H T +R
x(k|k) = x(k|k-1)
W = P(k|k-1) H T S -1
P(k|k) = P(k|k-1)
v(k) = z(k) - H x(k|k-1)
Figure 5.1: Flow chart for the Kalman filter. Note the recursive structure and how simple it
is to implement once the model matrices F and H have been specified along with the model
uncertainty matrices R and Q .
in distance-over-ground in normal flight conditions. Now comes the clever part. The
plane is equipped with an altimeter radar which measures height above the ground - a
direct measurement of z. Fusing this measurement in the Kalman filter will result in a
change in estimated height and also a change in y-position. The reason being that the
correlations between height and y-position maintained in the covariance matrix mean
that changes in estimated height should imply changes in estimated y-position because
the two states are co-dependent. The exercises associated with these lectures should
illustrate this fact to you further.
45
Using Estimation Theory in Mobile Robotics 46
Previous pages have contained a fair amount of maths taking you through the derivation of
various estimators in a steady fashion. Remember estimation theory is at the core of many
problems in mobile robotics — sensors are uncertain, models are always incomplete. Tools
like the Kalman filter 6 give a powerful framework in which we can progress.
We will now apply these freshly derived techniques to problems in mobile robotics fo-
cussing particularly on navigation. The following sections are not just a stand-alone-example.
Within the analysis, implementation and discussion a few new ideas are introduced which
will be used later in the course.
A Lander is at an altitude x above the planet and uses a time of flight radar to detect
altitude—see Figure 5.2.1. The onboard flight controller needs estimates of both height
and velocity to actuate control surfaces and time rocket burns. The task is to estimate both
altitude and descent velocity using only the radar. We begin modelling by assuming that the
vehicle has reached its terminal velocity (but this velocity may change with height slowly).
A simple model would be:
· ¸ · δT 2 ¸
1 δT
x(k) = x(k − 1) + 2 v(k) (5.58)
0 1 δT
| {z } | {z }
F G
where δT is the time between epochs and the state vector x is composed of two elements
altitude and rate of altitude change (velocity)
· ¸
h
x(k) = . (5.59)
ḣ
The process noise vector is a scalar, gaussian process with covariance Q. It represents noise
in acceleration (hence the quadratic time dependence when adding to a position-state).
Now we need to write down the observation equations. Think of the observation model
as “explaining the observations as a function of the thing you want to estimate”.
6
it is not the only method though
46
Using Estimation Theory in Mobile Robotics 47
Figure 5.2: A simple linear navigation problem. A Lander is at an altitude x above the
planet and uses a time of flight radar to detect altitude. The onboard flight controller needs
estimates of both height and velocity to actuate control surfaces and time rocket burns. The
task is to estimate both altitude and descent velocity using only the radar.
So we can write
These are the “truth models” of the system (note there are no “hats” or (k—k) notations
involved). We will never known the actual value of the noises at any epoch but we model the
imperfections in the sensor and motions models that they represent by using their covariance
matrices (R and Q respectively) in the filter equations.
We are now in a position to implement this example in Matlab. Section 11.1 is a print
out of the entire source code for a solution - it can also be downloaded from the course
website ¡http://www.robots.ox.ac.uk/∼pnewman : teaching¿. The exact source created the
following graphs and so all parameters can be read from the listing. It’s important that
you can reconcile all the estimation equations we have derived with the source and also that
you understand the structure of the algorithm —the interaction of simulator, controller and
estimator.
47
Using Estimation Theory in Mobile Robotics 48
8000
5
Altitude (m)
6000
0
4000
−5
2000
−10
100 200 300 400 500 0 200 400 600
1
0
Velocity (km/h)
0.5
−50
0
−100 −0.5
−1
−150
100 200 300 400 500 100 200 300 400 500
time (s) time (s)
The simulation model is non-linear. This of course is realistic — we may model the world
within our filter as a well behaved linear system 7 but the physics of the real world can be
relied upon to conspire to yield something far more complicated. However, the details of
the models employed in the simulation are not important. One point of this example is to
illustrate that sometimes we can get away with simple approximations to complex systems.
We only examine the vertical descent velocity (i.e. we ignore the coriolis acceleration)—
it’s as though we were dropping a weight vertically into a layer of atmosphere. The drag
exerted on the vehicle by the atmosphere is a function of both vehicle form-factor, velocity
and atmospheric density as a function of height (modelled as a saturating exponential).
7
we shall shortly introduce a way to use non-linear models but the idea that the world is always more
complicated than we care to imagine is still valid
48
Using Estimation Theory in Mobile Robotics 49
−5 Observations
x 10
0
50 100 150 200 250 300 350 400 450 500 550
time
−7 Innovation Sequence
x 10
Innovation (time of flight seconds)
−5
0 100 200 300 400 500 600
time
Figure 5.3:
Vehicle Controller
The controller implemented does two things. Firstly when the vehicle descends below a
predetermined height it deploys a parachute which increases its effective aerodynamic drag.
Secondly, it fires retro-burners when it drops below a second altitude threshold. A simple
control law based on altitude and velocity is used to control the vehicle to touch down with
a low velocity. At the point when rockets are fired, the parachute is also jettisoned.
The important point here is that the controller operates on estimated quantities. If the
estimated quantities are in error it is quite easy to envision bad things happening. This is a
point common to all robotic systems—actions (involving substantial energies) are frequently
executed on the basis estimates. The motivation to understand estimation process and its
failure modes is clear!
Flight Pattern Figure 5.2.1 shows the simulated(true) and estimated states using the code
listed in Section 11.1. Initially the vehicle is high in thin atmosphere which produces
little drag. The vehicle accelerates through the high levels of the atmosphere. Soon the
density increases and the vehicle brakes under the effect of drag to reach a quasi-steady
49
Using Estimation Theory in Mobile Robotics 50
state terminal velocity. When the parachute opens the instantaneous increase drag
decelerates the vehicle rapidly until another steady state terminal velocity is achieved.
Finally shortly before landing the retro-burners are fired to achieve a smooth landing
- essentially decelerating the vehicle smoothly until touch down.
Fitness of Model The filter vehicle model < F, G > is one of constant velocity (noise
in acceleration). Off-the-blocks then we should expect good estimation performance
during periods of constant velocity. Examining the graphs we see this is true. When
the velocity is constant both position and velocity are tracked well. However during
periods of rapid acceleration we see poor estimates of velocity emerging. Note that
during these times the innovation sequence and truth-estimated graphs ‘spike’....
Derivation of Velocity Note that the observation model H does not involve the velocity
state and yet the filter is clearly estimating and tracking velocity pretty well. At no
point in the code can you find statements like vel = (xnew − xold)/∆T . The filter
is not explicitly differentiating position to derive velocity — instead it is inferring it
through the models provided. The mechanism behind this has already been discussed
in section 5.1.3. The filter is using correlations (off diagonals) in the 2 × 2 matrix P
between position and velocity states. The covariance matrix starts off being diagonal
but during the first prediction step it becomes fully populated. 8 . Errors in position
are correlated through the vehicle model to errors in velocity. This is easy to spot in
the plant model as predicted position is a function of current position estimate and
velocity estimate. Here the KF is working as an observer of a hidden state — an
immensely useful characteristic. However there is no free lunch. Note how during
times of acceleration the velocity estimate lags behind the true velocity. This makes
sense 9 the velocity state is being dragged (and hence lags) through state space by the
correlations to directly observed position.
Filter Tuning An obvious question to ask is how can the filter be made “tighter”? How can
we produce a more agile tracker of velocity? The answer lies in part with the process
noise strength Q . The addition of GQGT at each time step dilutes the interstate
correlations. By making Q smaller we maintain stronger correlations and track inferred
velocity. But we cannot reduce Q too far— it has to model the uncertainty in our
model. If we reduce it too much we will have too much confidence in our predictions
and the update stage will have little corrective effect. The process of choosing a suitable
Q (and R ) is called tuning. It is an important part of KF deployment and can be hard
to do in practice. Fortunately there are a few concepts that can help in this process.
Their derivation is more suited for a course in stochastic estimation rather than mobile
8
You should download the code and check this out for yourself. Try forcing the off diagonals to zero
and see the effect
9
expand the KF update equations with a constant velocity model and full P matrix. Note how the
change in velocity W(2, 1) is a function of the off diagonals
50
Using Estimation Theory in Mobile Robotics 51
robotics and so some of them are stated here as a set of rules (valid for linear Gaussian
problems);
There is some skill involved in choosing values of R and Q such that the above crite-
ria are met, especially when the filter models are a poor representation of the truth.
The correct thing to do here is implement a better model. If however, other engi-
neering issues impede this course of action, the filter must be de-tuned (increase noise
strengths) in the hope of ‘lumping’ un-modelled characteristics into the noise vector.
This of course means that the filter looses any claim to optimality.
10
There is a powerful geometric interpretation of the Kalman filter that fits closely to this analogy using
ideas of orthogonality
51
Incorporating Non-Linear Models - The Extended Kalman Filter 52
Up until now we have only considered linear models (although working in non-linear simu-
lated environments). It shouldn’t come as a surprise to you that the majority of real world
applications require the use of non-linear models. Think about an everyday example - a re-
ally simple GPS receiver sitting at xr = (x, y, z)T and measuring time of flight (equivalent)
to a satellite sitting at xs = (x, y, z)Ts . Clearly the time of flight measurement is “explained”
by the norm of the difference vector || xr − xs . This then requires the use of a non-linear
measurement model. Fortunately we shall see that non-linear models can easily be incorpo-
rated into the Kalman Filter equations (yielding a filter called the Extended Kalman Filter
or EKF) you are already familiar with. The derivations are given here for completeness and
the results are stated in Section 5.3.3.
The trick behind the EKF is to linearize the non-linear models around the “best” current
estimate (best meaning prediction (k|k − 1) or last estimate (k − 1|k − 1)). This is done
using a Taylor series expansion. Assume we have an estimate x̂(k − 1|k − 1) then
x(k) = f (x̂(k − 1|k − 1), u(k), k) + ∇Fx [x(k − 1) − x̂(k − 1|k − 1)] + · · · (5.64)
52
Incorporating Non-Linear Models - The Extended Kalman Filter 53
Now we know that x̂(k|k − 1) = E{x(k)|Zk−1 } and x̂(k − 1|k − 1) = E{x(k − 1)|Zk−1 }so
x̂(k|k − 1) = E{f (x̂(k − 1|k − 1), u(k), k) + ∇Fx [x(k − 1) − x̂(k − 1|k − 1)] + · · · + v(k)|Zk−1 }
(5.65)
= E{f (x̂(k − 1|k − 1), u(k), k)|Zk−1 } + ∇Fx [x̂(k − 1|k − 1) − x̂(k − 1|k − 1)]
(5.66)
= f (x̂(k − 1|k − 1), u(k), k) (5.67)
Which is an obvious conclusion — simply pass the last estimate through the non-linear model
to come up with a prediction based on a control signal u (k). To figure out how to propagate
the covariance (using only terms from the previous time step) we look at the behaviour of
x̃(k|k − 1) = x(k) − x̂(k|k − 1) because P(k|k − 1) = E{x̃(k|k − 1)x̃(k|k − 1)T |Zk−1 }.
Understanding that the jacobians of f are evaluated at x̂(k − 1|k − 1) we can write
So
We now have the predict equations in the case of non-linear plant models. Note that fre-
quently the model will be in the form
where the noise v (k) is not simply additive. In this case one would proceed with a multivari-
ate Taylor11 series which swiftly becomes notationally complex and algebraically tiresome.
However the end result is intuitive. The state prediction remains unchanged but the predic-
tion equation becomes:
53
Incorporating Non-Linear Models - The Extended Kalman Filter 54
where ∇Fu is the jacobian of f w.r.t the noise vector. (Don’t worry many examples to follow
— also look at the source code provided ) It is a common practice to make the substitution
u(k) = un (k) + v(k) where un (k) is a nominal control which is then corrupted by noise. In
this case ∇Gv = ∇Gu . You will see this again soon. In the meantime, look at the example
source code provided.
Now we turn to the case of a non-linear observation model (for example a range and bearing
sensor) of the general form
z(k) = h(x(k)) + w(k) (5.80)
By using the same technique used for the non-linear prediction step we can show that the
predicted observation z(k|k − 1) (Hx̂(k|k − 1) in the linear case) is simply the projection of
x̂(k|k − 1) through the measurement model:
Now we wish to derive an expression for S, the innovation covariance. We begin by expressing
the observation and the estimated observation error z̃(k|k − 1) using a Taylor series:
So the covariance of the difference between actual and predicted observations (the innovation)
can be written as:
You may be wondering where the E{x̃(k|k − 1)w(k)T |Zk−1 } terms have gone. They evaluate
to zero as (reasonably) we do not expect the noise in observations to be correlated to the
error in our prediction.
54
Incorporating Non-Linear Models - The Extended Kalman Filter 55
We now have one last thing to figure out — how does a non-linear observation model
change the update equations resulting in x̂(k|k) and P(k|k)? The procedure should now be
becoming familiar to you: figure out an expression using a series expansion for x̃(k|k) and
take squared conditional expectation to evaluate P(k|k).
Assume that “somehow” we have a gain W (we’ll derive this in a minute) then we can
immediately write:
W ( z(k) − h(x̂(k|k − 1))
x̂(k|k) = x̂(k|k − 1) + |{z} (5.92)
| {z } |{z} | {z }
prediction gain observation predictedobservation
| {z }
innovation
We can now use this expression to make progress in figuring out P(k|k).
P(k|k) = E{x̃(k|k)x̃(k|k)T |Zk } (5.93)
x̃(k|k) = x̂(k|k) − x(k) (5.94)
= x̂(k|k − 1) + W(z(k) − h(x̂(k|k − 1))) − x(k) (5.95)
= x̂(k|k − 1) + Wz̃(k|k − 1) − x(k) (5.96)
55
Incorporating Non-Linear Models - The Extended Kalman Filter 56
where
Note that substituting P(k|k − 1)∇Hx T = WS into Equation 5.103 leads to the form of
the covariance update we are already familiar with:
We can now state in a “good to remember this box” a rule for converting the linear Kalman
filter equations to the non-linear form:
To convert the linear Kalman Filter to the Extended Kalman Filter simply replace
F with ∇Fx and H with ∇Hx in the covariance and gain calculations only.
The jacobians are always evaluated at the best available estimate (x̂(k − 1|k − 1)
for ∇Fx and x̂(k|k − 1) for ∇Hx
For completeness here are the EKF equations. (You’ll need these for the class-work). If
you don’t feel you are on top of the previous maths - its not the end of the world. Make
56
Incorporating Non-Linear Models - The Extended Kalman Filter 57
sure however you are familiar with the general form of the equations. (exam useful things
are in boxes in these notes (but not exclusively))
Prediction:
plant model
z }| {
x̂(k|k − 1) = f (x̂(k − 1|k − 1), u(k) , k) (5.111)
| {z } | {z } |{z}
predicted state old state est control
Update:
where
measurement
z}|{
ν(k) = z(k) −z(k|k − 1) (5.116)
T −1
W = P(k|k − 1)∇Hx S (5.117)
| {z }
kalman gain
57
Incorporating Non-Linear Models - The Extended Kalman Filter 58
We are now in a great position, possessing some very powerful tools which we shall now
apply to some key autonomous navigation problems.
58
Topic 6
This is a good point to write down a simple motion model for a mobile robotic vehicle. We
allow the vehicle to move on a 2D surface ( a floor) and point in arbitrary directions. We
parameterise the vehicle pose xv (the joint of position and orientation) as
xv
x v = yv . (6.1)
θv
Figure 6.1 is a diagram of a non-holonomic (local degrees of freedom less than global
degree of freedom) vehicle with “Ackerman” steering. The angle of the steering wheels is
given by φ and the instantaneous forward velocity (sometimes called throttle) is V . With
reference to Figure 6.1, we immediately see that
Using the instantaneous center of rotation we can calculate the rate of change of orien-
59
Velocity Steer Model 60
Instantaneous φ
Center of
Rotation φ
a V
(x,y) L
Note that we have started to lump the throttle and steer into a control vector — this makes
sense if you think about the controlling actions of a human driver. Equation 6.9 is a model for
a perfect, noiseless vehicle. Clearly this is a little unrealistic — we need to model uncertainty.
One popular way to do this is to insert noise terms into the control signal u such that
where un (k) is a nominal (intended) control signal and v(k) is a zero mean gaussian dis-
60
Evolution of Uncertainty 61
This completes a simple probabilistic model of a vehicle. We shall now see how propa-
gation of this model affects uncertainty in vehicle pose over time.
In this section we will examine how an initial uncertainty in vehicle pose increases over time
as the vehicle moves when only the control signal u is available. Hopefully you will realise
that one way to approach this is repetitive application of the prediction step of a Kalman
filter discussed in Section 5.1.1. The model derived in the previous section is non-linear and
so we will have to use the non-linear form of the prediction step.
Assume at time k we have been given a previous best estimate of the vehicle pose x̂v (k −
1|k − 1) and an associated covariance Pv (k − 1|k − 1). Equations 5.67 and 5.79 have that:
xv (k|k − 1) = f (x̂(k − 1|k − 1), u(k), k) (6.13)
T T
Pv (k|k − 1) = ∇Fx Pv (k − 1|k − 1)∇Fx + ∇Fv Q∇Fv (6.14)
In this case · ¸
σ2 0
Q= V (6.15)
0 σφ2
We need to evaluate the Jacobians with respect to state and control noise at x̂(k − 1|k − 1).
We do this by differentiating each row of f by each state and each control respectively:
1 0 −δT V sin(θv )
∇Fx = 0 1 δT V cos(θv ) (6.16)
0 0 1
δT cos(θv ) 0
∇Fu = δT sin(θv ) 0 (6.17)
tan(φ) δT V sec2 (φ)
L L
Figure 6.2 shows the results of iterating equations 6.13 and 6.14 using the matlab code
printed in Section 11.2. Things are pretty much as we might expect. The uncertainty
61
Evolution of Uncertainty 62
20
15
y 10
−5
−20 −15 −10 −5 0 5 10 15 20
x
Figure 6.2: Evolution of uncertainty evolution of open loop the Ackermann Model. The
circles are the true location of the vehicle whereas the crosses mark the dead-reckoned lo-
cations. The orientation of the vehicle is made clear by the orientation of the triangles.
Note the divergence between true and dead-reckoned locations. This is typical of all dead
reckoning methods — the only thing that can be changed is the rate of divergence.
injected into the system via the noisy control makes the estimated covariance of the vehicles
grow without bound.
There is an important point to make here that you must understand. In actual real life
the real robot is integrating the noisy control signal. The true trajectory will therefore always
drift away from the trajectory estimated by the algorithms running inside the robot. This is
exactly the same as closing your eyes and trying to walk across University Parks. Your inner
ears and legs give you u which you pass through your own kinematic model of your body
in your head. Of course, one would expect a gross accumulation of error as the time spent
walking “open loop” increases. The point is that all measurements such as velocity and
rate of turn are measured in the vehicle frame and must be integrated, along with the noise
on the measurements. This always leads to what is called “dead reckoning drift”. Figure
6.3 shows the effect of integrating odometry on a real robot called “B21” shown in figure
6.3(right). The main cause of this divergence on land vehicles is wheel slip. Typically robot
wheels are fitted with encoders that measure the rotation of each wheel. Position is then
an integral-function of these “wheel counts”. The problem is a wheel or radius r may have
turned through θ but due to slip/skid the distance travelled over the ground is only (1 − η)rθ
where η is an unobservable slip parameter.
62
Using Dead-Reckoned Odometry Measurements 63
The model in Section 6.1 used velocity and steer angles as control input into the model. It is
common to find that this low level knowledge is not easy to obtain or that the relationship
between control, prior and prediction is not readily discernable. The architecture in figure
6.3 is a typical example.
Hidden
Control
Physics
Other Measurements
Encoder Counts
Nav
DR Model
Figure 6.3: Sometimes a navigation system will be given a dead reckoned position as input
without recourse to the control signals that were involved. Nevertheless the dead-reckoned
position can be converted into a control input (a stream of small motions) for use in the core
navigation system.
The low level controller on the vehicle reads encoders on the vehicle’s wheels and outputs
an estimate (with no metric of uncertainty) of its location. We can make a guess at the
kind of model it uses1 . Assume it has two wheels (left and right), radius r mounted either
side of its center of mass which in one time interval turn an amount δθl , δθr — as shown in
Figure 6.3. We align a body-centered co-ordinate frame on the vehicle as shown. We want
1
This for illustration only - the real b21 vehicle is actually a synchronous drive machine in which all four
wheels change direction http://www.irobot.com
63
Using Dead-Reckoned Odometry Measurements 64
Figure 6.4: Dead Reckoned position from a B21 mobile robot. The start and end locations
are actually the same place! See how you could roll the trajectory back over itself. This is
typical of dead reckoned trajectories — small angular errors integrate to give massive long
term errors
to express the change of position of the center of the vehicle as a function of δθl , δθr :
The dead-reckoning system in the vehicle simply compounds these small changes in position
64
Using Dead-Reckoned Odometry Measurements 65
dθ y'
x'
]
,dy
y
[dx
α
r
L x
c
and orientation to obtain a global position estimate. Starting from an initial nominal frame
at each iteration of its sensing loop it deduces a small change in position and orientation,
and then “adds” this to its last dead-reckoned position. Of course the “addition” is slightly
more complex than simple adding (otherwise the x coordinate would always be zero!). What
actually happens is that the vehicle composes successive co-ordinate transformation. This
is an important concept (which you will probably have met in other courses but perhaps
with a different notation) and will be discussed in the next section.
Figure 6.3.1 shows three relationships between three coordinate frames. We can express any
coordinate j frame with respect to another frame i as a three-vector xi,j = [xyθ]. Here x and
y are translations in frame i to a point p and θ is anti-clockwise rotation around p. We define
two operators ⊕ and ª to allow us to compose (chain together) multiple transformations:
65
Using Dead-Reckoned Odometry Measurements 66
X 2,3
X 1,2
X 1,3
With reference to figure 6.3.1 we see that x1,3 = x1,2 ⊕ x2,3 . But what exactly are these
operators? Well, they are just a short hand for a function of one or two transformations:
· ¸
x + x2 cos θ1 − y2 sin θ1
x1 ⊕ x 2 = 1 (6.26)
y1 + x2 sin θ1 + y2 cos θ1
−x1 cos θ1 − y1 sin θ1
ªx1 = x1 sin θ1 − y1 cos θ1 (6.27)
−θ1
Be sure you understand exactly what these equations allow. They allow you to express
something (perhaps a point or vehicle) described in one frame, in another alternative frame.
We can use this notation to explain the compounding of odometry measurements. Figure
6.3.1 shows a vehicle with a prior pose xo (k − 1). The processing of wheel rotations between
successive readings (via Equation 6.23) has indicated a vehicle-relative transformation (i.e.
in the frame of the vehicle) u . The task of combining this new motion u(k) with the old
dead-reckoned estimate xo to arrive at a new dead-reckoned pose xo is trivial. It is simply:
We have now explained a way in which measurements of wheel rotations can be used to esti-
mate dead-reckoned pose. However we set out to figure out a way in which a dead-reckoned
pose could be used to form a control input or measurement into a navigation system. In
66
Using Dead-Reckoned Odometry Measurements 67
x o(k)
u(k)
x o (k-1)
Global Frame
other words we are given from the low-level vehicle software a sequence xo (1), xo (2) · · · xo (k)
etc and we want to figure out u (k). This is now simple— we can invert equation 6.28 to get
Looking at the Figure 6.3.1 we can see that the transformation u(k) is equivalent to going
back along xo (k − 1) and forward along xo (k). This gives us a small control vector u (k)
derived from two successive dead-reckoned poses that is suitable for use in another hopefully
less error prone navigation algorithm. Effectively equation 6.29 subtracts out the common
dead-reckoned gross error - locally odometry is good - globally it is poor.
We are now in a position to write down a plant model for a vehicle using a dead-reckoned
position as a control input:
It is reasonable to ask how an initial uncertainty in vehicle pose Pv propagates over time.
We know that one way to address this question is to propagate the second order statistics
(covariance) of a pdf for xv through f using equation 5.79. To do this we need to figure out the
jacobians of equation 6.32 with respect to xv and u. This is one area where the compositional
representation we have adopted simplifies matters. We can define and calculate the following
jacobians:
67
Using Dead-Reckoned Odometry Measurements 68
∂(x1 ⊕ x2 )
J1 (x1 , x2 ) , (6.33)
∂x1
1 0 −x2 sin θ1 − y2 cos θ1
= 0 1 x2 cos θ1 − y2 sin θ1 (6.34)
0 0 1
∂(x1 ⊕ x2 )
J2 (x1 , x2 ) , (6.35)
∂x2
cos θ1 − sin θ1 0
= sin θ1 cos θ1 0 (6.36)
0 0 1
Here the matrix Uo describes the strength of noise in the small shifts in pose represented by
uo derived from two sequential dead-reckoned poses. A simple form of this matrix would be
purely diagonal 2 : 2
σox 0 0
2
Uo = 0 σoy 0 (6.39)
2
0 0 σoθ
where the diagonals are variances in odometry noise. For example if the odometry loop ran
at 20Hz and the vehicle is moving at 1m/s the magnitude of translation in u would be 5cm.
If we say slip accounts for perhaps one percent of distance travelled we might “try” a value
2 2
of σox = σoy = (0.05/100)2 . Allowing a maximum rotation of w perhaps a good starting
guess for σoθ would be (w/100)2 . These numbers will give sensible answers while the vehicle
2
is moving but not when it is stopped. Even when uo = 0 the covariance Pv will continue
to inflate. This motivates the use of a time varying Uo which is a function of uo (k). An
exercise you should think about.....
2
implying we expect errors in x y and orientations to be uncorrelated which is probably not true in reality
but we will live with this approximation for now
68
Topic 7
7.1 Introduction
We will now apply the estimation techniques we have learnt to two very important mobile
robotics tasks - Mapping and Localisation. Theses two tasks are fundamental to successful
deployment of autonomous vehicles and systems for example:
A common way to approach these problems is to parameterise both the robot pose and
aspects of the environment’s geometry into one or more state vectors. For this course we
69
Features and Maps 70
will work mostly in 2D but the definitions that follow are, of course, valid for the full 3D
case.
We suppose that the world is populated by a set of discrete landmarks or features whose
location / orientation and geometry (with respect to a defined coordinate frame) can be
described by a set of parameters which we lump into a feature vector xf .
In this course we will constrain ourselves to using the simplest feature possible - a point
feature such that for the ith feature:
· ¸
x
xf ,i = i (7.2)
yi
where x and y are the coordinates of the point in a global frame of reference. Point features
are not as uncommon as one might initially think. Points occur at the intersection of lines,
corners of rectangles, edges of objects (regions of maximal curvature)1 .
7.3 Observations
We define two distinct types of observations all denoted as z — vehicle relative and absolute.
Absolute Absolute observations are made with the help of some external device and usually
involve a direct measurement of some aspect of the vehicle’s pose. The best examples
1
Sure, our own vision system operates at a higher level recognising things like gorillas and sail-boats as
complete entities but lower-level geometric-primitive detection is buried in there as well.
70
A Probabilistic Framework 71
are a GPS and a compass. They depend on external infrastructure (taken as known)
and are nothing to do with the map.
Vehicle Relative This kind of observation involves sensing the relationship between the
vehicle and its immediate surroundings –especially the map, see figure 7.3. A great
example is the measurement of the angle and distance to a point feature with respect to
the robot’s own frame of reference. We shall also class odometry (integration of wheel
movement) as a vehicle-relative measurement because it is not a direct measurement
of the vehicle’s state.
x fi
Mobile Vehicle
xv
It is informative to describe the localisation and mapping tasks in terms of likelihoods (ob-
servation pdf) and priors.
For the localisation task we assume we have been given a map and receive a sequence of
vehicle-relative observations described by a likelihood p(Zk |M, xv ). We wish to figure a pdf
71
A Probabilistic Framework 72
If we are given a nominally perfect map (i.e. told to take it as absolute truth) then p(M) = 1
this simplifies the above:
For the mapping side of things we are given the vehicle location (probably derived from
absolute observation) and an sequence of vehicle relative observations. We wish to find the
distribution of M conditioned on Zk and xv .
72
Feature Based Estimation for Mapping and Localising 73
Equations 7.15 and 7.8 should have a familiar form — we met them when discussing
Maximum a priori estimators, recursive bayesian estimation which collectively lead us to
discuss and explore the Kalman filter. The Kalman filter would appear to be an excellent
way in which to implement these equations. If we parameterise the random vectors x v and
M with first and second order statistics (mean and variance) then the Kalman Filter will
calculate the MMSE estimate of the posterior. The first derivation of the Kalman filter
presented proceeded by assuming Gaussian distributions. In this case the Kalman filter is
the optimal Bayesian estimator. The Kalman filter provides a real time way to perform state
estimation on board a vehicle.
This is the simplest task. We are given a map M containing a set of features and a stream
of observations of measurements between the vehicle and these features (see figure 7.3). We
assume to begin with that an oracle is telling us the associations between measurements and
observed features. We assume the vehicle we are navigating is equipped with a range-bearing
sensor which returns the range and bearing to point like objects (the features). We will base
the ensuing simulation on a real vehicle - the B21 we have already met - this means that
we will have an additional sequence of dead-reckoned positions as input into the prediction
stage. We denote these as xo
We have already have the prediction equations from previous discussions ( Equation
7.16):
73
Feature Based Estimation for Mapping and Localising 74
Now we come to the observation equation which is simply a range ri and bearing θi to
the ith feature:
· ¸
r
z(k) , (7.20)
θ
= h(x(k), w(k)) (7.21)
"p #
(xi − xv (k))2 + (yi − yv (k))2
= (7.22)
atan2( xyii −y v (k)
−xv (k)
) − θv
We differentiate w.r.t xv to arrive at the observation model jacobian:
∂h
∇Hx , (7.23)
"∂x #
xi −xv (k) yi −yv (k)
r r
0
= yi −yv (k) (7.24)
r2
− xi −xr2v (k) −1
We assume independent errors on range and bearing measurements and use a diagonal
observation covariance matrix R : · 2 ¸
σr 0
R= (7.25)
0 σθ2
Section 11.3 is a print out of an implementation of feature based localisation using the
models we have just discussed. You can also download the code from the course web-site
¡http://www.robots.ox.ac.uk/∼pnewman : teaching¿. Figure 7.5.1 shows the trajectory of the
vehicle as it moves through a field of random point features. Note that the code simulates
a sensor failure for the middle twenty percent of the mission. During this time the vehicle
becomes more and more lost. When the sensor comes back on line there is a jump in
estimated vehicle location back to one close to the true position (see figure 7.5.1).
Now we will consider the dual of Localisation - Mapping. In this case the vehicle knows
where it is but not what is in the environment. Perhaps the vehicle is fitted with a GPS
74
Feature Based Estimation for Mapping and Localising 75
60
40
20
−20
−40
−60
−80 −60 −40 −20 0 20 40 60 80
20
5
0
x
range
0 −20
−40
−5 0 1000 2000 3000 4000 5000 6000
20
−10
0 1000 2000 3000 4000 5000 6000 10
0
y
30 −10
20 −20
0 1000 2000 3000 4000 5000 6000
Bearing (deg)
10 200
0 100
−10 0
θ
−20 −100
−30 −200
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000
time time
Figure 7.3: Feature Based Localisation innovation and error/covariance plots. Notice how
when no measurements are made the vehicle uncertainty grows rapidly but reduces when
features are re-observed. The covariance bounds are 3-sigma and 1-sigma-bound on state
error and innovation plots respectively.
or some other localisation system using an a-priori map. To avoid initial confusion we’ll
imagine the vehicle has a ’super-gps’ on board telling it where it is at all times.
The state vector for this problem is now much larger -it will be the Map itself and
75
Feature Based Estimation for Mapping and Localising 76
the concatenation of all point features. The observation equation is the same as for the
localisation case only now the feature co-ordinates are the free variables.
The prediction model for the state is trivial. We assume that features don’t move and
so x(k + 1|k)map = x(k|k)map .
vehicle trajectory
Figure 7.4: Evolution of a map over time. The vertical axis is time (k). Covariance ellipses
for each feature are plotted in the z = k planes. Note how the filter converges to a perfect
map because no process noise is added in the prediction step
Initially the map is empty and so we need some method to add “newly discovered’ features
to it. To this end we introduce a feature initialisation function Y that take as arguments
the old state vector and an observation to a landmark and returns a new, longer state vector
with the new feature at its end. For the case of a range bearing measurement we would have
the following:
76
Feature Based Estimation for Mapping and Localising 77
where the coordinates of the new feature are given by the function g:
We also need to figure out how to transform the covariance matrix P when adding a new
feature. Of course we can use the jacobian of the transformation:
· ¸
P(k|k) 0
∗
P(k|k) = ∇Yx,z ∇Yx,z T (7.32)
0 R
where
· ¸
I 0n×2
∇Yx,z = n×n (7.33)
∇Gx ∇Gz
Now for the mapping case ∇Gx = 0 as the new feature is not dependent on any element
in the state vector (which only contains features). Therefore:
· ¸
∗ P(k|k) 0
P(k|k) = (7.34)
0 ∇Gz R∇Gz T
One thing to realise is that the observation Jacobian is now “long and thin”. When
observing feature i it is only non-zero at the “location” (indexes) of the feature in the state
vector:
· ¸
∇Hx = |· · ·{z
0 · ·}· ∇Hxfi
| {z } |· · ·{z
0 · ·}·
(7.35)
otherf eatures observedf eature otherf eatures
Figure 7.5.2 shows the evolution of the map over time. Note that as no process noise is
ever added to the system the uncertainty in feature locations after initialisation is always
decreasing. In the limit the map will be known perfectly. The code used to generate this
simulation for feature based mapping can be downloaded from the course web-site. You
might think it odd that in the limit the map becomes perfectly known. You might think
that intuitively there should always be some residual uncertainty. The point is that the
77
Simultaneous Localisation and Mapping - SLAM 78
vehicle is continually being told where (via the super-GPS) it is and is not changing its
position estimate as a function of landmark observations. As a consequence all features
are independent (check this by examining the P matrix –it is block diagonal) and each
observation of them simply adds information and therefore reduces their uncertainty - again
and again and again.... This is in contrast to the next section where landmark observations
will be allowed to adjust map and vehicle estimates simultaneously!
SLAM is the generalised navigation problem. It asks if it is possible for a robot, starting
with no prior information, to move through its environment and build a consistent map of
the entire environment. Additionally the vehicle must be able to use the map to navigate
(localise) and hence plan and control its trajectory during the mapping process. The appli-
cations of a robot capable of navigating , with no prior map, are diverse indeed. Domains
in which ’man in the loop’ systems are impractical or difficult such as Martian exploration,
sub-sea surveys, and disaster zones are obvious candidates. Beyond these, the sheer increase
in autonomy that would result from reliable, robust navigation in large dynamic environ-
ments is simply enormous. Autonomous navigation has been an active area of research for
many years.
In SLAM no use is made of prior maps or external infrastructure such as GPS. A SLAM
algorithm builds a consistent estimate of both environment and vehicle trajectory using only
noisy proprioceptive (e.g., inertial, odometric) and vehicle-centric (e.g., radar, camera and
laser) sensors. Importantly, even though the relative observations are of the local environ-
ment, they are fused to create an estimate of the complete workspace.
The SLAM problem is of fundamental importance in the quest for autonomous mobile
machines. It binds aspects of machine learning, uncertainty management, perception, sensing
and control to enable a machine to discover and understand its surroundings with no prior
knowledge or external assistance.
This section will introduce a simple feature based approach to the SLAM problem. It
doesn’t work in real life for deployments in large areas because it involves running a Kalman
filter to estimate the entire map and vehicle state. The update of the covariance matrix
is therefore at best proportional to the square of the number of features. Given enough
features the system will grind to a halt. Figure 7.6 shows some of the kind of area that can
be mapped and localised in using this technique.
78
Simultaneous Localisation and Mapping - SLAM 79
Figure 7.5: SLAM in a real environment. These figures are screen shots from an exper-
iment using a B21 robot in a hallway at MIT. Having built the map the vehicle used
it to navigate home to within a few cm of its start position. Videos can be found at
¡http://www.robots.ox.ac.uk/∼pnewman : teaching¿.
You’ll be pleased to know we have pretty much done all the work required to implement
a full SLAM algorithm. We will still employ the oracle that tells us which feature is being
seen with each observation. All we do now is change our state vector to include both vehicle
and map:
xv
x , xf ,1
(7.36)
..
.xf ,n
Of course to start with n = 0 but as new features are seen the state vector is grown
79
Simultaneous Localisation and Mapping - SLAM 80
just as in the pure mapping case. The difference now is that the observation and feature
initialisation jacobians have two non-zero blocks. one with respect to the vehicle and one
with respect to the observed feature.
£ ¤
∇Hx = ∇Hxv · · · ∇Hxf ,1 (7.37)
· ¸
In×n 0n×2
∇Yx,z = (7.38)
∇Gx ∇Gz
· ¸
In×n 0n×2
= £ ¤ (7.39)
∇Gxv · · · 0 · · · ∇Gz
Also note that the jacobian of the prediction models have changed in an obvious way:
· ¸
∇Fxv 0
∇Fx = (7.40)
0 I2n×2n
· ¸
J1(xv , u) 0
= (7.41)
0 I2n×2n
· ¸
J2(xv , u) 0
∇Fu = (7.42)
0 02n×2n
(7.43)
Note that the features are not expected to move between time steps (pretty much what you
would hope for when using a map!) and they are noiseless. Only the vehicle has process
noise injected into its covariance matrix during a prediction step.
The matlab code for an EKF, feature based SLAM algorithm can be found on the course
web-site. You should be able to recognise it as a union of previous localisation and mapping
examples. Figure 7.6 shows a few snap shots of the algorithm running.
80
Simultaneous Localisation and Mapping - SLAM 81
80 80 80
60 60 60
40 40 40
20 20 20
0 0 0
80 80 80
60 60 60
40 40 40
20 20 20
0 0 0
80 80 80
60 60 60
40 40 40
20 20 20
0 0 0
Note that the covariance matrix now has some structure to it -you can partition map P mm
and vehicle Pvv blocks.
· ¸
Pvv Pvm
P= (7.45)
PTvm Pmm
Of diagonals Pvm are the correlations between map and vehicle. Its pretty clear that there
should be correlations between map and vehicle as they are so interrelated. Consider the
following sequence of events. From the moment of initialisation the feature’s location is a
function of vehicle location and so errors in the vehicle location will also appear as errors
in feature location. Now recall the discussion regarding correlations in section 5.1.3 —
correlations cause adjustments in one state to ripple into adjustments in other states. This
is of course also true in this kind of approach to the SLAM problem. Remarkably every
observation of a feature affects the estimate of every other feature in the map. It’s as though
they are all tied up together with elastic bands - pulling at one will pull at the others in
turn.
81
Simultaneous Localisation and Mapping - SLAM 82
There are some further characteristics of the SLAM problem that transcend the estima-
tion method being used. You should be able to check that they are true by running the
example code:
• The feature uncertainties never drop below the initial uncertainty of the vehicle. This
makes sense, if you start with say 10m of uncertainty relative to Carfax tower mapping
and re-observing things within the Thom building is never going to reduce your overall
uncertainty. The best you can hope for is to reduce all features to the lower bound -
your initial uncertainty. Compare this to the ever decreasing uncertainty of features
in the mapping only case.
• The feature uncertainties never increase but the vehicle uncertainty can. The prediction
model is the identity matrix for the map - we don’t expect it to move. Furthermore the
map has a noiseless model and so the prediction step does not inflate the covariance
of the map.
The SLAM problem is a very topical problem and is attracting interest from the AI,
robotics and machine learning communities. If you want to find out more just ask....
82
Topic 8
The previous discussion on Estimation, Mapping and Localisation has focused almost exclu-
sively on estimating uni-model pdfs. Much use has been made of the Kalman filter which
estimates the mean and covariance of a pdf which under gaussian noise assumptions is the
MMSE estimator. The Kalman filter is an important tool (and has great value outside of
this course) but it is not the whole story.
The single EKF approach we have talked about so far does not really behave well in
large environments - the big problem is that linearisation errors (that came from the Taylor
series approximations in deriving the EKF) compound and the linearisation starts to occur
around points that are far from the true state. There is neither the space nor time to cover in
detail how this can be overcome. Broadly though, one of two schemes are adopted. The first
partitions the world into local maps and uses uni-modal methods (EKF) within each map.
The maps are then carefully glued back together to form a unified global map. Secondly
uni-modal methods are abandoned altogether. Instead, monte-carlo methods are used.
The basic idea is simple. We maintain lots of different versions of the state vector — all
slightly different from each other. When a measurement comes in we score how well each
version explains the data. We then make copies of the best fitting vectors and randomly
perturb them (the equivalent of adding Q in the EKF) to form a new generation of candidate
states. Collectively these thousands of possible states and their scores define the pdf we are
seeking to estimate. We never have to make the assumption of Gaussian noise or perform
83
Montecarlo Methods - Particle Filters 84
a linearisation. The big problem with this elegant approach is the number of sample-states
(versions) we need to try out increase geometrically with the number of states being esti-
mated. For example to estimate a 1D vector we may keep 100 samples. For a 2D vector
we would require something like 1002 , for 3D 1003 and so on. The trick often played is to
be smart about selecting which samples should be copied between timesteps. This is a very
current area of research.
2. Apply the plant model to each particle. In contrast to the prediction step in the
Kalman approach we actually inject the process noise as well — i.e we add a random
vector to the control vector u. This is in some ways analogous to adding GUGT to
the predicted covariance in the Kalman approach because it increases the variance of
the estimate (in this case the particle set).
3. For each particle predict the observation. Compare this to the measured value. Use a
likelihood function to evaluate the likelihood of the observation given the state repre-
sented by each particle. This will be a scalar L which is associated with each particle.
This scalar is referred to as a “weight” or “importance”.
4. Select the particles that best explain the observation. One way to do this would be to
simply sort and select the top candidates based on L but this would reduce the number
of particles. One solution to this would be to copy from this “winning set” until we
have n particles again. This has a hidden consequence though: it would artificially
reduce the diversity (spread) of the particle set. Instead we randomly sample from the
samples biased by the importance values. This means that there is a finite chance that
particles that really didn’t explain the observation well, will be reselected. This may
turn out to be a good plan because they may do a better job for the next observation.
5. Goto 2
Figure 8.1 shows this process graphically for the localisation problem. A common ques-
tion is “what is the estimate?”. Well technically it is the distribution of particles themselves
which represents a pdf. If you want a single vector estimate, the mode, mean and median are
all viable options. However you need to think carefully. If you run the code yourself initially
you will see distinct clouds of particles representing differing modes of the pdf - the mean
is somewhere between them all in a location with no support1 ! As with all the algorithms
1
however soon the clouds should coverge into one large cloud as the motion of the vehicle disambiguates
its location
84
Grid Based Mapping 85
in the course you are strongly advised to down load the code (MCL.m) and peruse it to
understand the translation from concept to implementation. In summary, the particle filter
approach is very elegant and very simple to implement (see the code on the website). A
crucial point is that it does not require any linearisation assumptions (there are no jacobians
involved) and there are no Gaussian assumptions. It is particularly well suited for problems
with small state spaces - the example code has a state dimension of three.
Up until now we have adopted a feature based approach to navigation. We assumed that the
environment contains features possessing some geometry that can be extracted by suitable
processing of sensor data - edges, points, lines etc. We then store and estimate (using further
measurements) this sparse parameterisation (often by stacking all the feature parameters in
to a long state vector and calling it x ).
In contrast grid based approaches tend not to reason about the geometry of features in
the environment but take a more sensor-centric view. A mesh of cells (a grid) is laid over the
(initially unknown) world. Each cell c(i, j) has a number associated with it corresponding
to an occupancy probability - po (i, j) and a probability of being empty pe . So for example
if po (40, 2) = 1 we are sure that the cell at (40,20) is occupied. We make no assertions as to
what is occupying it just that something or part of something is at the coordinates covered
by cell(40,20).
We use two functions to model the behaviour of the sensor - one to express the likelihood
of a cells within the observed region being empty and one to express the likelihood of them
being occupied. Figure 8.2 shows one possible set of functions for a widebeam in-air sonar.
For each cell under the beam a Bayes rule update is used to figure out the new probability
of a cell being empty or being occupied. As the sensor (mounted on a vehicle) moves around
the environment and if the location of the sensor is known a grid based map can be built.
Figure 8.2 shows the occupancy and empty maps for a map built with the CMU Terragator
vehicle in figure 8.2.
One problem with grid based methods is memory usage - it becomes expensive for 3D
maps. However cunning data structures can overcome this. A second issue is that of grid
alignment - it is arbitrary and as such cells are likely to cover areas that are awkwardly
partially full as such is not possible to capture the sharp discontinuities at the edges of
objects. Perhaps the biggest problem of grid based methods stems from the difficulties in
localising with respect to a grid. It requires trying to match a small sensor centric map with
85
Grid Based Mapping 86
1. Plant Model
x i = f(x i,u(k),v(k)) 1 motion +
perturbation
5
1
5 2
2 6
4 6
starting sample set 3 4
7
7
3
Predicted Observations
1 Feature
5
Actual Observation 2
6
4
h(x i) 2. Obs Prediction
7
3
p(z|x i )
3. Fitness Evaluation
L i = f(z-h(x i))
4 3 21 5 6 7
z-h(x i ) obs-pred
cumulative importance ( ΣL i)
uniformly
choosen
random 4. Selection
numbers
particle 1 2 3 4 5 6 7
#copies 1 1 3 1 1 0 0
1
5. Regeneration
5
updated sample set 2
4
333
x3
Figure 8.1: Graphical interpretation of a particle filter. The method is general but this
diagram can be easily understood in terms of a range-to-beacon sensor (like the long baseline
acoustic sensor mentioned in the Least Squares section). The best particles are those that
best explain the actual range measurement. These are preferentially choosen to form the
next generation.
86
Grid Based Mapping 87
Figure 8.2: From the early days of mobile robotics. The CMU terragator vehicle used grid
based mapping in various settings.
Figure 8.3: The shape of two functions used to describe the probability Pe (r) of empty cells
at a range r from the sensor and the probability of them being occupied po (r). The actual
object is at distance R from the sensor. The sensor modelled here is a wide beam sonar that
sends out a cone of sound and measures the time between transmit and echo reception.
87
Grid Based Mapping 88
Figure 8.4: The two typical occupancy maps built by a grid based approach. The left map
show the probability of regions being occupied and the left the probability of regions being
empty. Each map is maintained using a different likelihood function
a large global map. The vehicle can be localised by rotating and translating the local map
(what is seen by the sensor) until the best match is found.
88
Topic 9
In Conclusion
Hopefully you have found this introductory course interesting and have picked up a sense
that the tools we have been using are applicable to many domains. If this were a 10 lecture
course we would have spent more time considering how to appraise in real-time the validity
of sensor data. This is a crucial point — we need to decide whether the data is even worth
processing. In may situations sensors return grossly erroneous data (outliers). However if
we don’t know what the world looks like how can we be sure it is bad data and not faithful
sensing of an unexpected world state? I also hope you have found the discussions on mobile
robotics thought provoking. Perhaps in a decade we will have cars that avoid crashes, goods
trucks that drive themselves, an always-deployed fleet of ocean vehicles monitoring El-Nino
currents, smart machines on Mars and robots that massively increase the independence and
mobility of our old and infirm. We don’t need the “full-monty” of AI to be able to make a
big difference....(but it would be useful).
89
Topic 10
Miscellaneous Matters
This section is meant to help you with the class work and covers things that you may need
to know but that don’t really fit in the main body of the notes
1 1
p(x) = exp{− (x − µx )T P−1 (x − µx )} (10.1)
(2π)n/2|P| 1/2 2
xT P−1 x = 1 (10.2)
90
Drawing Covariance Ellipses 91
p(x,y)
Figure 10.1: A bi-variate normal distribution. The thick contour is the 1 − σ bound where
(x − µx )T P−1 (x − µx ) = 1
.
xT VD−1 VT x = 1 (10.4)
T −1/2 −1/2 T
x VD D V x=1 (10.5)
xT KKT x = 1 (10.6)
where
K = VD−1/2 . (10.7)
now for any point y = [x, y]T which is on the unit circle, yT y = 1, so
xT KKT x = yT y (10.8)
T
K x=y (10.9)
1/2
⇒ x = VD y (10.10)
2σ xx
2σ yy
Figure 10.2: The relationship between geometry of the 1 − σ bound for a bivariate normal
distribution and its covariance matrix.
Figure 10.1 shows the relation ship between P and the plotted form for a bivariate normal
distribution.
· ¸
2
σ 2 σxy
P = xx2 2 (10.11)
σxy σyy
2
2σxy
tan 2α = 2 2
(10.12)
σxx − σyy
92
Drawing High Dimensional Gaussians 93
Obviously it is hard to draw a high dimensional gaussian on a screen or paper. For example
we may have a covariance matrix that corresponds to a large state vector:
a
x = b (10.13)
c
· ¸¾
σa11 2 σa12 2
σa21 2 σa22 2 Paa ··· ···
marginal of component b
z }| {
· ¸¾
P= .. σb11 2 σb12 2 (10.14)
. 2 2 Pbb ···
σb21 σb22 · ¸¾
.. .. σc11 2 σc12 2
. . 2 2 Pcc
σc21 σc22
So to plot a 2D representation of the ith entity in a state vector simply plot the ellipse for
the ith 2 by 2 block in P .
93
Topic 11
Example Code
%set up parameters
DoInitialise ;
k = 2;
while(˜VehicleStatus.Landed & k <Params.StopTime/Params.dT)
94
Matlab Code For Mars Lander Example 95
% store results
DoStore(k,XEst,PEst,Innovation,S,z(k));
%tick ...
k = k+1;
end;
return;
95
Matlab Code For Mars Lander Example 96
VehicleStatus.Landed = 0;
%observation expected noise strength (we never know this parameter exactly)
%set the scale factor to 1 to make model and reallity match
Params.R = (1.1∗Params.SigmaR)ˆ2;
Params
return;
%prediction ...
XPred = F∗XEst;
PPred = F∗PEst∗F’+G∗Q∗G’;
96
Matlab Code For Mars Lander Example 97
% do update....
XEst = XPred+W∗Innovation;
PEst = PPred−W∗S∗W’;
return;
oldvel = XTrue(2,k−1);
oldpos = XTrue(1,k−1);
dT = Params.dT;
function DoControl(k,XEst)
if (XEst(1)<Params.RocketBurnHeight )
if (˜VehicleStatus.RocketsOn)
fprintf(’Releasing Chute at time %f\n’,k∗Params.dT);
fprintf(’ Firing Rockets at time %f\n’,k∗Params.dT);
Params.BurnTime = k∗Params.dT;
end;
%turn on thrusters
VehicleStatus.RocketsOn = 1;
97
Matlab Code For Ackerman Model Example 98
%drop chute..
VehicleStatus.Drag = 0;
%simple littel controller here (from vˆ2 = uˆ2+2as) and +mg for weight of vehicle
VehicleStatus.Thrust = (Params.m∗XEst(2)ˆ2−1)/(2∗XEst(1))+0.99∗Params.m∗Params.g;
end;
if (XEst(1)<1)
%stop when we hit the ground...
fprintf(’Landed at time %f\n’,k∗Params.dT);
VehicleStatus.Landed = 1;
Params.LandingTime = k∗Params.dT;
break;
end;
return;
else
Store.XEst = [Store.XEst XEst];
Store.PEst = [Store.PEst diag(PEst)];
Store.Innovation = [ Store.Innovation Innovation];
Store.S = [Store.S diag(S)];
Store.z = [Store.z z ];
end;
return;
Q = diag([SigmaVˆ2 SigmaPhiˆ2]);
98
Matlab Code For Ackerman Model Example 99
xlabel(’x’ ); ylabel(’y’ );
title ( ’uncertainty bounds for Ackermann model’);
%calculate jacobians
JacFx = [1 0 −dT∗u(1)∗sin(x(3)); 0 1 dT∗u(1)∗cos(x(3)); 0 0 1];
JacFu = [dT∗cos(x(3)) 0; dT∗sin(x(3)) 0; dT∗ tan(u(2))/L dT ∗u(1)∗sec(u(2))ˆ2];
%prediction steps
P = JacFx ∗ P ∗JacFx’ + JacFu∗Q∗JacFu’;
xtrue = AckermannModel(xtrue,u+[SigmaV ;SigmaPhi].∗randn(2,1),dT,L);
x = AckermannModel(x,u,dT,L);
%draw occasionally
if (mod(k−1,30)==0)
PlotEllipse (x,P ,0.5); DrawRobot(x,’r’);plot(xtrue(1),xtrue(2), ’ko’ );
end;
end;
99
Matlab Code For EKF Localisation Example 100
nSteps = 6000;
Map = 140∗rand(2,30)−70;
UTrue = diag([0.01,0.01,1∗pi/180]).ˆ2;
RTrue = diag([2.0,3∗pi/180]).ˆ2;
UEst = 1.0∗UTrue;
REst = 1.0∗RTrue;
xTrue = [1;−40;−pi/2];
xOdomLast = GetOdometry(1);
% initial conditions :
xEst =xTrue;
PEst = diag([1,1,(1∗pi/180)ˆ2]);
% initial graphics
figure (1); hold on; grid off ; axis equal;
plot(Map(1,:),Map(2,:),’g∗’ ); hold on;
set(gcf, ’doublebuffer’ , ’on’ );
hObsLine = line ([0,0],[0,0]);
set(hObsLine,’ linestyle ’ , ’ :’ );
for k = 2:nSteps
100
Matlab Code For EKF Localisation Example 101
u = tcomp(tinv(xOdomLast),xOdomNow);
xOdomLast = xOdomNow;
%do prediction
xPred = tcomp(xEst,u);
xPred(3) = AngleWrap(xPred(3));
PPred = J1(xEst,u)∗ PEst ∗J1(xEst,u)’ + J2(xEst,u)∗ UEst ∗ J2(xEst,u)’;
if (˜isempty(z))
%predict observation
zPred = DoObservationModel(xPred,iFeature,Map);
S = jH∗PPred∗jH’+REst;
W = PPred∗jH’∗inv(S);
xEst = xPred+ W∗Innov;
xEst(3) = AngleWrap(xEst(3));
else
%Ther was no observation available
xEst = xPred;
PEst = PPred;
Innov = [NaN;NaN];
S = NaN∗eye(2);
end;
if (mod(k−2,300)==0)
DoVehicleGraphics(xEst,PEst (1:2,1:2),8,[0,1]);
if (˜isempty(z))
set(hObsLine,’XData’,[xEst(1),Map(1,iFeature)]);
set(hObsLine,’YData’,[xEst(2),Map(2,iFeature)]);
end;
drawnow;
end;
%store results :
InnovStore(:,k) = Innov;
PStore(:,k) = sqrt(diag(PEst));
SStore (:, k) = sqrt(diag(S));
XStore(:,k) = xEst;
XErrStore(:,k) = xTrue−xEst;
101
Matlab Code For EKF Localisation Example 102
end;
DoGraphs(InnovStore,PStore,SStore,XStore,XErrStore);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function DoGraphs(InnovStore,PStore,SStore,XStore,XErrStore)
figure(2);
subplot(2,1,1);plot(InnovStore (1,:)); hold on;plot(SStore(1,:), ’ r ’ ); plot(−SStore(1,:),’ r ’ )
title ( ’Innovation’ ); ylabel(’range’);
subplot(2,1,2);plot(InnovStore(2,:)∗180/pi);hold on;plot(SStore(2,:)∗180/pi,’r’ ); plot(−SStore(2,:)∗180/pi,’r’)
ylabel(’Bearing (deg)’);xlabel(’time’);
print −depsc ’EKFLocationInnov.eps’
figure(2);
subplot(3,1,1);plot(XErrStore(1,:));hold on;plot(3∗PStore(1,:),’r ’ ); plot(−3∗PStore(1,:),’r’ );
title ( ’Covariance and Error’);ylabel(’x’);
subplot(3,1,2);plot(XErrStore(2,:));hold on;plot(3∗PStore(2,:),’r ’ ); plot(−3∗PStore(2,:),’r’ )
ylabel(’y’ );
subplot(3,1,3);plot(XErrStore(3,:)∗180/pi);hold on;plot(3∗PStore(3,:)∗180/pi,’r’);plot(−3∗PStore(3,:)∗180/pi,’r’)
ylabel(’\theta’ ); xlabel(’time’);
print −depsc ’EKFLocationErr.eps’
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [z,iFeature] = GetObservation(k)
global Map;global xTrue;global RTrue;global nSteps;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [z] = DoObservationModel(xVeh, iFeature,Map)
Delta = Map(1:2,iFeature)−xVeh(1:2);
z = [norm(Delta);
atan2(Delta(2),Delta(1))−xVeh(3)];
z(2) = AngleWrap(z(2));
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function SimulateWorld(k)
global xTrue;
u = GetRobotControl(k);
xTrue = tcomp(xTrue,u);
xTrue(3) = AngleWrap(xTrue(3));
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
102
Matlab Code For EKF Mapping Example 103
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [xnow] = GetOdometry(k)
persistent LastOdom; %internal to robot low−level controller
global UTrue;
if (isempty(LastOdom))
global xTrue;
LastOdom = xTrue;
end;
u = GetRobotControl(k);
xnow =tcomp(LastOdom,u);
uNoise = sqrt(UTrue)∗randn(3,1);
xnow = tcomp(xnow,uNoise);
LastOdom = xnow;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function u = GetRobotControl(k)
global nSteps;
u = [0; 0.025 ; 0.1∗ pi/180∗sin(3∗pi∗k/nSteps)];
%u = [0; 0.15 ; 0.3∗ pi /180];
nSteps = 600;
nFeatures = 6;
MapSize = 200;
Map = MapSize∗rand(2,nFeatures)−MapSize/2;
UTrue = diag([0.01,0.01,1∗pi/180]).ˆ2;
RTrue = diag([8.0,7∗pi/180]).ˆ2;
UEst = 1.0∗UTrue;
REst = 1.0∗RTrue;
xVehicleTrue = [1;−40;−pi/2];
103
Matlab Code For EKF Mapping Example 104
%storage:
PStore = NaN∗zeros(nFeatures,nSteps);
XErrStore = NaN∗zeros(nFeatures,nSteps);
for k = 2:nSteps
if (˜isempty(z))
S = jH∗PPred∗jH’+REst;
W = PPred∗jH’∗inv(S);
xEst = xPred+ W∗Innov;
PEst = PPred−W∗S∗W’;
%note use of ’Joseph’ form which is numerically stable
% I = eye(size (PEst));
104
Matlab Code For EKF Mapping Example 105
PEst = M∗blkdiag(PEst,REst)∗M’;
%remember this feature as being mapped we store its ID and position in the state vector
MappedFeatures(iFeature,:) = [length(xEst)−1, length(xEst)];
end;
else
%There was no observation available
end;
if (mod(k−2,40)==0)
plot(xVehicleTrue(1),xVehicleTrue(2),’r∗’ );
fprintf(’k = %d\n’,k);
drawnow;
end;
%Storage:
for(i = 1:nFeatures)
if (˜isnan(MappedFeatures(i,1)))
iL =MappedFeatures(i,1);
PStore(k,i) = det(PEst(iL:iL+1,iL:iL+1));
XErrStore(k,i) = norm(xEst(iL:iL+1)−Map(:,i));
end;
end;
end;
figure(2);
plot(PStore);
105
Matlab Code For EKF Mapping Example 106
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [z,iFeature] = GetObservation(k)
global Map;global xVehicleTrue;global RTrue;global nSteps;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [z] = DoObservationModel(xVeh, xFeature)
Delta = xFeature−xVeh(1:2);
z = [norm(Delta);
atan2(Delta(2),Delta(1))−xVeh(3)];
z(2) = AngleWrap(z(2));
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function SimulateWorld(k)
global xVehicleTrue;
u = GetRobotControl(k);
xVehicleTrue = tcomp(xVehicleTrue,u);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [jHxv,jHxf] = GetObsJacs(xPred, xFeature)
jHxv = zeros(2,3);jHxf = zeros(2,2);
Delta = (xFeature−xPred(1:2));
r = norm(Delta);
jHxv(1,1) = −Delta(1) / r;
jHxv(1,2) = −Delta(2) / r;
jHxv(2,1) = Delta(2) / (rˆ2);
jHxv(2,2) = −Delta(1) / (rˆ2);
jHxv(2,3) = −1;
jHxf(1:2,1:2) = −jHxv (1:2,1:2);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [jGx,jGz] = GetNewFeatureJacs(Xv, z);
x = Xv(1,1);
y = Xv(2,1);
theta = Xv(3,1);
r = z(1);
bearing = z(2);
jGx = [ 1 0 −r∗sin(theta + bearing);
0 1 r∗cos(theta + bearing)];
jGz = [ cos(theta + bearing) −r∗sin(theta + bearing);
sin(theta + bearing) r∗cos(theta + bearing)];
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function u = GetRobotControl(k)
global nSteps;
%u = [0; 0.25 ; 0.3∗ pi/180∗sin(3∗pi∗k/nSteps)];
u = [0; 0.15 ; 0.3∗ pi/180];
106
Matlab Code For EKF SLAM Example 107
%length of experiment
nSteps = 8000;
%size of problem
nFeatures = 40;
MapSize = 200;
Map = MapSize∗rand(2,nFeatures)−MapSize/2;
UTrue = diag([0.01,0.01,1.5∗pi/180]).ˆ2;
RTrue = diag([1.1,5∗pi/180]).ˆ2;
UEst = 2.0∗UTrue;
REst = 2.0∗RTrue;
xVehicleTrue = [0;0;−pi/2];
%storage:
PStore = NaN∗zeros(nFeatures,nSteps);
XErrStore = NaN∗zeros(nFeatures,nSteps);
xOdomLast = GetOdometry(1);
107
Matlab Code For EKF SLAM Example 108
for k = 2:nSteps
xPred = [xVehiclePred;xMap];
PPred = [PPredvv PPredvm;
PPredvm’ PPredmm];
if (˜isempty(z))
%have we seen this feature before?
if ( ˜isnan(MappedFeatures(iFeature,1)))
zPred = DoObservationModel(xVehicle,xFeature);
S = jH∗PPred∗jH’+REst;
W = PPred∗jH’∗inv(S);
xEst = xPred+ W∗Innov;
108
Matlab Code For EKF SLAM Example 109
PEst = PPred−W∗S∗W’;
PEst = M∗blkdiag(PEst,REst)∗M’;
%remember this feature as being mapped we store its ID and position in the state vector
MappedFeatures(iFeature,:) = [length(xEst)−1, length(xEst)];
end;
else
xEst = xPred;
PESt = PPred;
end;
if (mod(k−2,DrawEveryNFrames)==0)
a = axis;
clf ;
axis(a);hold on;
n = length(xEst);
nF = (n−3)/2;
DoVehicleGraphics(xEst(1:3),PEst (1:3,1:3),3,[0 1]);
if (˜isnan(z))
h = line([xEst(1),xFeature (1)],[ xEst(2),xFeature (2)]);
set(h, ’ linestyle ’ , ’ :’ );
end;
for(i = 1:nF)
iF = 3+2∗i−1;
plot(xEst(iF),xEst(iF+1),’b∗’);
PlotEllipse (xEst(iF:iF+1),PEst(iF:iF+1,iF:iF+1),3);
end;
fprintf(’k = %d\n’,k);
drawnow;
end;
if (ismember(k,SnapShots))
iPic = find(SnapShots==k);
print(gcf,’−depsc’,sprintf(’EKFSLAM%d.eps’,iPic));
end;
end;
109
Matlab Code For EKF SLAM Example 110
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [z,iFeature] = GetObservation(k)
global Map;global xVehicleTrue;global RTrue;global nSteps;global SensorSettings
done = 0;
Trys = 1;
z =[]; iFeature = −1;
while(˜done & Trys <0.5∗size(Map,2))
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [z] = DoObservationModel(xVeh, xFeature)
Delta = xFeature−xVeh(1:2);
z = [norm(Delta);
atan2(Delta(2),Delta(1))−xVeh(3)];
z(2) = AngleWrap(z(2));
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function SimulateWorld(k)
global xVehicleTrue;
u = GetRobotControl(k);
xVehicleTrue = tcomp(xVehicleTrue,u);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [jHxv,jHxf] = GetObsJacs(xPred, xFeature)
jHxv = zeros(2,3);jHxf = zeros(2,2);
Delta = (xFeature−xPred(1:2));
r = norm(Delta);
jHxv(1,1) = −Delta(1) / r;
jHxv(1,2) = −Delta(2) / r;
jHxv(2,1) = Delta(2) / (rˆ2);
jHxv(2,2) = −Delta(1) / (rˆ2);
jHxv(2,3) = −1;
jHxf(1:2,1:2) = −jHxv (1:2,1:2);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [jGx,jGz] = GetNewFeatureJacs(Xv, z);
x = Xv(1,1);
y = Xv(2,1);
theta = Xv(3,1);
r = z(1);
bearing = z(2);
jGx = [ 1 0 −r∗sin(theta + bearing);
110
Matlab Code For Particle Filter Example 111
0 1 r∗cos(theta + bearing)];
jGz = [ cos(theta + bearing) −r∗sin(theta + bearing);
sin(theta + bearing) r∗cos(theta + bearing)];
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [xnow] = GetOdometry(k)
persistent LastOdom; %internal to robot low−level controller
global UTrue;
if (isempty(LastOdom))
global xVehicleTrue;
LastOdom = xVehicleTrue;
end;
u = GetRobotControl(k);
xnow =tcomp(LastOdom,u);
uNoise = sqrt(UTrue)∗randn(3,1);
xnow = tcomp(xnow,uNoise);
LastOdom = xnow;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function u = GetRobotControl(k)
global nSteps;
%u = [0; 0.25 ; 0.3∗ pi/180∗sin(3∗pi∗k/nSteps)];
u = [0; 0.15 ; 0.2∗ pi/180];
function MCL
close all ; clear all ;
global xTrue;global Map;global RTrue;global UTrue;global nSteps;
nSteps = 6000;
111
Matlab Code For Particle Filter Example 112
nParticles = 400;
Map = 140∗rand(2,30)−70;
UTrue = diag([0.01,0.01,1∗pi/180]).ˆ2;
RTrue = diag([2.0,3∗pi/180]).ˆ2;
UEst = 1.0∗UTrue;
REst = 1.0∗RTrue;
xTrue = [1;−40;−pi/2];
xOdomLast = GetOdometry(1);
% initial graphics
figure (1); hold on; grid off ; axis equal;
plot(Map(1,:),Map(2,:),’g∗’ ); hold on;
set(gcf, ’doublebuffer’ , ’on’ );
hObsLine = line ([0,0],[0,0]);
set(hObsLine,’ linestyle ’ , ’ :’ );
hPoints = plot(xP(1,:),xP (2,:), ’ .’ );
for k = 2:nSteps
%do prediction
%for each particle we add in control vector AND noise
%the control noise adds diversity within the generation
for(p = 1:nParticles)
xP(:,p) = tcomp(xP(:,p),u+sqrt(UEst)∗randn(3,1));
end;
xP(3,:) = AngleWrap(xP(3,:));
if (˜isempty(z))
112
Matlab Code For Particle Filter Example 113
%predict observation
for(p = 1:nParticles)
%what do we expect observation to be for this particle ?
zPred = DoObservationModel(xP(:,p),iFeature,Map);
%how different
Innov = z−zPred;
%get likelihood (new importance). Assume gaussian here but any pdf works!
%if predicted obs is very different from actual obs this score will be low
%−>this particle is not very good at representing state . A lower score means
%it is less likely to be selected for the next generation ...
L(p) = exp(−0.5∗Innov’∗inv(REst)∗Innov)+0.001;
end;
end;
if (mod(k−2,10)==0)
figure(1);
set(hPoints,’XData’,xP(1,:));
set(hPoints,’YData’,xP(2,:));
if (˜isempty(z))
set(hObsLine,’XData’,[xEst(1),Map(1,iFeature)]);
set(hObsLine,’YData’,[xEst(2),Map(2,iFeature)]);
end;
figure(2);plot(xP(1,:),xP (2,:), ’ .’ );
drawnow;
end;
end;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [z,iFeature] = GetObservation(k)
global Map;global xTrue;global RTrue;global nSteps;
113
Matlab Code For Particle Filter Example 114
z = [];
iFeature = −1;
else
iFeature = ceil(size(Map,2)∗rand(1));
z = DoObservationModel(xTrue, iFeature,Map)+sqrt(RTrue)∗randn(2,1);
z(2) = AngleWrap(z(2));
end;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [z] = DoObservationModel(xVeh, iFeature,Map)
Delta = Map(1:2,iFeature)−xVeh(1:2);
z = [norm(Delta);
atan2(Delta(2),Delta(1))−xVeh(3)];
z(2) = AngleWrap(z(2));
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function SimulateWorld(k)
global xTrue;
u = GetRobotControl(k);
xTrue = tcomp(xTrue,u);
xTrue(3) = AngleWrap(xTrue(3));
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [xnow] = GetOdometry(k)
persistent LastOdom; %internal to robot low−level controller
global UTrue;
if (isempty(LastOdom))
global xTrue;
LastOdom = xTrue;
end;
u = GetRobotControl(k);
xnow =tcomp(LastOdom,u);
uNoise = sqrt(UTrue)∗randn(3,1);
xnow = tcomp(xnow,uNoise);
LastOdom = xnow;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function u = GetRobotControl(k)
global nSteps;
u = [0; 0.025 ; 0.1∗ pi/180∗sin(3∗pi∗k/nSteps)];
114