A Primer on Odometry and Motor Control

Edwin Olson
[email protected]
December 4, 2004

1 Introduction
Most robotics problems ultimately reduce to the question: where am I? How can you go
somewhere else without some notion of where you are now? You could proceed naively,
stumbling around, hoping that your goal will appear in range of your robot’s sensors, but
a better plan is to navigate. A basic method of navigation, used by virtually all robots, is
odometry, using knowledge of your wheel’s motion to estimate your vehicle’s motion.
The goal of this primer is to show you how to use odometry data. A real dataset will be
used extensively to show how well (and how badly) these ideas can work.
Suppose your robot starts at the origin, pointed down the x-axis. Its state is (x, y, θ) =
(0, 0, 0). If the robot travels (roughly) straight for three seconds at 1 m/s, a good guess of
your state will be (3, 0, 0). That’s the essence of odometry.
We’ll assume that the vehicle is differentially driven: it has a motor on the left side of
the robot, and another motor on the right side. If both motors rotate forward, the robot
goes (roughly) straight. If the right motor turns faster than the left motor, the robot will
move left.
Our goal is to measure how fast our left and right motors are turning. From this, we
can measure our velocity and rate of turn, and then integrate these quantities to obtain our
position.

2 Derivation
The speeds of our motors give us two quantities: the rate at which the vehicle is turning,
and the rate at which the vehicle is moving forward. All we have to do is integrate these
two quantities, and we’ll have our robot’s state (x, y, θ).
That sounds a bit scary, but the mathematics end up being very simple. If we had
analytic expressions for the angular rate and forward rate functions, their integral probably
would be scary. But in a real system, our data comes from real sensors that are sampled
periodically. Every few milliseconds, we get a new measurement of our motor velocities. We
will numerically integrate our solution, which is of course, just a fancy way of saying that
we’ll divide time up into little chunks and just add up all the little pieces.
Suppose our robot is at (x, y, θ). Depending on what kind of sensors we have, we might
get measurements of how much the motors have rotated (in radians), or an estimate of how
fast they’re rotating (angular rate, radians/sec.) We’ll consider the first case, but of course
the second case can be handled by just multiplying the angular rate by the amount of time
that has elapsed since our last iteration. Given the amount of rotation of the motor and
the diameter of the wheel, we can compute the actual distance that the wheel has covered.

1
 Suppose the left wheel has moved by a distance of dlef t and the right wheel has moved
dright . For some small period of time (such that dlef t and dright are short), we can reasonably
assume that the vehicle trajectory was an arc (see Fig. 1).

θ'

d r ig
ht
(x',y')
d lef

θ
t

(x,y)
ϕ
dbaseline
P

Figure 1: Odometry geometry. Over a small time period, the robot’s motion can be ap-
proximated by an arc. The odometry problem is to solve for (x0 , y 0 , θ0 ) given (x, y, θ) and
dbaseline . In the figure, the robot is moving counter-clockwise.

The initial state (x, y, θ) defines our starting point, with θ representing the vehicle’s
heading. After our vehicle has moved by dlef t and dright , we want to compute the new
position, (x0 , y 0 , θ0 ).
The center of the robot (the spot immediately between the two wheels that define’s the
robot’s location), travels along an arc as well. Remembering that arc length is equal to the
radius times the inner angle, the length of this arc is:
dlef t + dright
dcenter = (1)
2
Given basic geometry, we know that:

φrlef t = dlef t (2)

φrright = dright (3)

If dbaseline is the distance between the left and right wheels, we can write:

rlef t + dbaseline = rright (4)

Subtracting (2) from (3), we see:

φrright − φrlef t = dright − dlef t

φ(rright − rlef t ) = dright − dlef t

2
 θ'

d r ig
ht
(x',y')

d ce
nt
er
d lef

θ
t

(x,y)
ϕ

P θ−π/2
rleft
rcenter
rright

Figure 2: Detailed odometry geometry. We compute the center of the circle P by forcing
dlef t and dright to be two arcs with the same inner angle φ. From this, we can compute the
new robot position (x0 , y 0 , θ0 ).

φdbaseline = dright − dlef t

dright − dlef t
φ= (5)
dbaseline
At the risk of making the math a bit uglier than really necessary, let’s very carefully
compute the new robot state. All of our arcs have a common origin at point P . Note that
the angle of the robot’s baseline with respect to the x axis is θ − π/2. We now compute the
coordinates of P :

Px = x − rcenter cos(θ − π/2)

= x − rcenter sin(θ)
Py = y − rcenter sin(θ − π/2)
= y + rcenter cos(θ)

Now we can compute x0 and y 0 :

x0 = Px + rcenter cos(φ + θ − π/2)

= x − rcenter sin(θ) + rcenter ∗ sin(φ + θ)
= x + rcenter [−sin(θ) + sin(φ)cos(θ) + sin(θ)cos(φ)] (6)

and

3
 y0 = Py + rcenter sin(φ + θ − π/2)
= y + rcenter cos(θ) − rcenter cos(φ + θ)
= y + rcenter [cos(θ) − cos(φ)cos(θ) + sin(θ)sin(φ)] (7)

If φ is small (as is usually the case for small time steps), we can approximate sin(φ) = φ
and cos(φ) = 1. This now gives us:

x0 = x + rcenter [−sin(θ) + φcos(θ) + sin(θ)]

= x + rcenter φcos(θ)
= x + dcenter cos(θ) (8)

and

y0 = y + rcenter [cos(θ) − cos(θ) + φsin(θ)]

= y + rcenter [φsin(θ)]
= y + dcenter sin(θ) (9)

In summary, our odometry equations for (x0 , y 0 , θ0 ) reduce to:

dlef t + dright
dcenter = (10)
2
dright − dlef t
φ = (11)
dbaseline
θ0 = θ+φ (12)
0
x = x + dcenter cos(θ) (13)
y0 = y + dcenter sin(θ) (14)

3 Basic Motor Theory

Subsequent sections will discuss how we can measure the performance of our motors. We’ll
need to put this into context by talking first about some basic motor physics.
Motors are devices which manipulate magnetic fields in order to convert electrical energy
into rotational energy. An electric current flows through coils of wire, causing a magnetic
field to form. This magnetic field interacts with permanent magnets (like those on your
refrigerator), causing the shaft to rotate.
We also know that motors can be used in reverse; if you rotate the shaft of the motor,
the permanent magnets induce a current in the rotating coils of wire. This, of course, is
known as a generator.
It is important to realize that motors are always behaving as generators. If the output
shaft is rotating, a current is being induced in the wires. Usually, this current opposes the
current resulting from an externally applied voltage.

3.1 Motor model

A motor can be simply modelled by the resistance of its coil of wire and a voltage source that
produces a voltage proportional to rotational speed (see Fig. 3). The “generator” voltage

4
 +
VEMF
+ -
VApplied
-

Motor

Figure 3: Simple motor model. The simplest motor model is a resistor (corresponding to the
resistance of the motor’s wire coil), and a voltage source corresponding to the Back EMF.

is known as the Electro-Motive Force (EMF). It’s often called the “Back EMF”, since it’s
the voltage coming back out of the motor.
When we interface with a motor, we typically control the voltage applied to the motor.
This applied voltage can cause current to flow into the motor. At any given moment, Ohm’s
law can be applied. However, the voltage that we have to consider isn’t the applied voltage,
but the net voltage: Vapplied − Vemf . In other words, the current can be written:

I = (Vapplied − Vemf )/Rwinding (15)

The current of a motor is extremely important: it is proportional to the torque. A motor
only does work when its producing torque, and since torque is proportional to current, this
only occurs when Vapplied 6= Vemf .
When a voltage is first applied to a motor, Vapplied − Vemf > 0, and so current will flow.
This will produce torque and the output shaft will begin to rotate (provided the torque
exceeds the static friction of the load.) As time goes on, the motor will slowly accelerate.
As the motor gains speed, Vemf increases, and the current drops. At some point, we’ll reach
an equilibrium, where the motor’s torque (and current) exactly matches the load.
Here’s an important thought experiment: suppose we take a motor and short the two
leads together. Will it be easier or harder to turn the output shaft? If we turn the output
shaft, the motor will act as a generator, producing a voltage of Vemf . But the applied
voltage is zero, since the leads are shorted. All of the back EMF voltage is applied across
the motor winding, inducing a current. This current will generate a torque that will oppose
the external torque. Setting both terminals of a motor to the same voltage brakes the motor,
without the need for any mechanical mechanism.

3.2 Applied Voltage and PWM

As we’ve seen, the applied voltage is intimately involved in the behavior of a motor. Special
circuitry is required to drive motors; a typical microcontroller can only provide a very small
amount of power on each of its outputs– far too little to power a motor. We’ll need some
additional circuitry to drive the motors.
It turns out that building a circuit that can output an arbitrary applied voltage at high
current is quite difficult. It is much easier to build a simple high current switch, whose
output is either Vmax or zero. In fact, we can build a switch that can change state very fast,
upwards of 100kHz.

5
 Suppose you’re trying to build a dimmable light switch; can you build one from just
an on-off switch? Suppose you start flipping the light on and off repeatedly, at first fairly
slowly. The room is bright, then dark, then bright again. Not a very good dimmer! But you
can probably imagine that if you could flip the switch fast enough, that you could achieve
intermediate brightness levels. This is what we’re going to do with motors.
If the switch is “on” 50% of the time, you might expect the motor to run about the same
as it does at 50% the voltage. This does, in fact, work... provided that the switching is fast
enough. The criteria for the switching rate is that the voltage has to change faster than the
motor can react; the motor has an “electrical time constant”, which reflects how rapidly it
can react to changing electrical conditions.

period

20%

50%

80%

Figure 4: PWM Waveforms. We can approximate different applied voltages by using a

pulse-width-modulated signal whose period is less than the motor’s electrical time constant.

This technique is called Pulse Width Modulation (PWM) (see Fig. 4), since the width
of each pulse is modulated in order to approximate different values of Vapplied .
In general, motor electrical time constants are on the order of a few milliseconds, but
PWM frequencies in the ¡20kHz range can cause audible vibrations. So often, PWM fre-
quencies are increased well above the human range of hearing.

3.3 Deficiencies of our model

This is a pretty good model for a motor, but it ignores one often important detail: electrical
transients. When the motor is in steady state (i.e., it’s been rotating under a uniform load
for a long time), our model is accurate. But when conditions are changing (the load is
changing, or the applied voltage is changing), the model can be pretty lousy.
The main issue we’ve overlooked is the transient behavior of the motor. Not only is
the coil of wire an electromagnet, it’s also an inductor. Recall that the voltage across an
inductor with inductance L is:
dI
V =L (16)
dt
If the current is changing quickly, the inductor can develop a very large voltage! Many
motor controllers support a controllable “slew rate”; this is a way of limiting the rate
at which the applied voltage changes (and hopefully the current as well!) Large voltage
transients are not good for circuits, and large motors can develope enormous transients on
the order of tens or hundreds of volts.

3.4 Summary
We’ll continue to use our simple model for the rest of this article. Here’s what you need to
remember:

• The Back EMF of a motor (Vemf ) is proportional to the speed of the motor.

6
 • Motor current is proportional to torque.
• We drive motors with PWM, which allows us to approximately control Vapplied .

4 Dead-reckoning Sensors
In order to build an odometry system, you must have a way of measuring the angular
velocity of the motor. This can be accomplished in a number of ways:
• Mono Phase Encoders. A sensor is placed on the motor shaft or wheel such that
when the shaft rotates, the circuitry generates alternating 1’s and 0’s. One way of
implementing a simple encoder is by attaching a disc with holes to the shaft, and using
a breakbeam sensor to detect when a hole passes by. A mono phase encoder can not
determine the direction of motion. A serious failure mode occurs when the motor is
essentially stationary with a hole half-way in front of the sensor: environmental noise
can cause the encoder to trigger and untrigger, making it appear as though the shaft
is rotating.

Break-beam sensor

Clockwise rotation:

Counter-Clockwise rotation:

Figure 5: Simple Encoder. As a perforated disk rotates, a break-beam sensor alternates

between 1 and 0. The direction of motion is ambiguous.

• Quadrature Phase Encoders. An improvement upon mono phase encoders, quadrature

phase encoders use two simple encoders arranged so that they produce waveforms 90
degrees out-of-phase. When the shaft rotates in one direction, signal A will “lead”
signal B; when rotating in the opposite direction, signal “B” will “lead” signal A.
Quadrature phase encoders are generally considered the most precise form of motor
feedback, and can be quite expensive. A simple state machine can decode the quadra-
ture phase signals into “up” and “down” pulses. The two out-of-phase signals provide
immunity from spurious transitions due to a hole being half-way in front of one of
the sensors: only one signal can be transitioning at any given time, and if this signal
jumps back and forth, it maps to alternating (and cancelling) forward and backward
motion.
• Tachometer. A tachometer is essentially a small motor attached to the output shaft of
the main motor. It operates as a generator, where the output voltage is proportional
to the angular velocity of the main motor.
• Direct Back EMF measurement. With a little extra circuitry, we can measure the back
EMF of a motor directly. We set Vapplied = 0, wait for the electrical transients to die
off, and then measure the voltage produced by the motor. Since we’re not applying
any power to the motor, the motor will tend to slow down while we’re performing the
measurement. Fortunately, it doesn’t take long for electrical transients to die off; we
can measure the back EMF quickly, and then restore power.

7
 Break-beam sensors
a
(90 degrees offset)
b
Clockwise rotation:
a
b

Counter-Clockwise rotation:
a
b

Figure 6: Quadrature Phase Encoder. Two simple encoders, arranged so that they pick up
the same perforated holes, but 90 degrees out of phase with each other, allow the direction
of rotation to be determined.

• Indirect Back EMF estimation via current sense. We can measure the amount of
current flowing through the motor by putting a small-valued resistor in the path
between the motor and ground. Ohm’s law tells us that the voltage across that
resistor will be proportional to the current.
If we know the motor current, the applied voltage (which we’re controlling), and the
resistance of the motor winding, we can estimate the back EMF. Solving Eqn 15 for
Vemf

Vemf = Vapplied − IRwinding (17)

• Gyro heading adjustment. Some vehicles have a gyroscope, which can provide an
estimate of the vehicle’s angular rate dθ/dt. Gyros drift over time, but can be quite
accurate over small periods of time. They can be used to improve dead reckoning
accuracy, particularly when a poor motor feedback sensor is used (e.g., back EMF.)

5 Odometry in Practice
Odometry can provide good dead-reckoning over short distances, but error accumulates very
rapidly. At every step, we inject error (noise) into not just x and y, but also θ. The error
in θ is the killer, since every error in θ will be amplified by future iterations.
There are a number of basic noise sources.

• Sensor error. If we’re using anything other than quadrature phase encoders for our
feedback, then our estimates of the dlef t and dright will be noisy.
• Slippage. When the robot turns a corner, one wheel will, in general, slip a little bit.
This means that even if the odometry data was perfect, the robot’s path would not
recoverable from it.
• Error in estimate of dbaseline or in wheel diameters. We must measure the baseline
of the robot and the scale factor relating angular rate and linear velocity (a function
of the wheel diameter.) Even small errors in these constants can cause problems–
typically, the odometry result will have a consistent veer in one direction or the other.
In fact, often the wheels on the robot are slightly different sizes. Even small differences
lead to large navigation errors with time.

8
Figure 7: Splinter. This is the robot that collected the dataset. The vehicle, known as
“Splinter”, has high quality quadrature phase encoders on each drive motor. A laptop on
the back of the vehicle logged the data. The large blue device in the front of the vehicle is
a laser range finder, not used in this article.

You should be itching to try out odometry now, and fortunately, there’s a dataset waiting
for you! In this section, we’ll walk through a few examples in Matlab.
The dataset consists of 200 meters of odometry and gyro data from an experiment that
lasted about 10 minutes. The vehicle that collected the data is shown in Fig. 7. The vehicle
left the lab, travelled around a large open space in a clockwise (as viewed from the top)
direction three times, then reversed direction and travelled around the same loop once in
the opposite direction. Finally, it returned to lab.
A good odometry result is shown in Fig. 8. While the robot started and stopped in
(approximately) the same space, notice how the odometry estimate is quite different at the
beginning and end. A small error in θ causes all subsequent navigation to be rotated out of
alignment.
The dataset is available from /mit/6.186/2005/doc/odomtutorial. Copy it into a direc-
tory and start matlab. The data is contained in a single array called “data”. Each row
in the dataset is a set of observations made at a single point in time. A helper script,
“setup.m” is provided which loads the dataset and sets up a few constants. These constants
are TIME, LEFTTICKS, RIGHTTICKS, LEFTCURRENT, RIGHTCURRENT, LEFT-
PWM, RIGHTPWM, and GYRO, and correspond to the column numbers in which data of
different types is stored. To get started, type:

>> setup

Note that the odometry data is provided in units of ticks: the amount of forward (or
backward) motion that occured since the last sample. Ticks are related to distance (and
angular rate) by constant scale factors, which are characteristics of the robot’s geometry.
We mentioned before that the baseline, the distance between the left and right wheels, is
also a characteristic of the robot.
These measurements are made using a ruler, and are inexact. One of our tasks will be
to find better values of the parameters.
Parameter Value
baseline 0.45 meters
metersPerTick 0.000045 meters

9
 Corrected Odometry

16

14

12

10
Y (meters)

−15 −10 −5 0
X (meters)

Figure 8: Good odometry result. This result was obtained by manually optimizing the rigid
parameters of the vehicle until the path best resembled the actual trajectory. Note that
the trajectory, in fact, started and ended at approximately the same position; even good
odometry results can contain significant errors.

Problem 1. Examining the raw odometry data

Let’s get familiar with the data by plotting the left versus right wheel odometry ticks; we’ll
be able to see how fast the left and right wheels are moving.

>> clf;
>> hold on;
>> plot(data(:,LEFTTICKS),’r’);
>> plot(data(:,RIGHTTICKS),’b’);

The command “clf” clears the figure, “hold on” allows multiple plots to be displayed on
top of each other, and the two plot commands plot the vector of data in either red (’r’) or
blue(’b’).
Compare the instantaneous velocity of the left and right wheels to the odometry results
in Fig. 8. Identify in the velocity plot where the robot was turning.

Problem 2. Plotting the gyro data

Now let’s take a look at the gyro’s reported angular position over time. Plot the gyro data
over time (the commands should look much like a previous example.) Your figure should
look something like Fig. 9.
The shape of this graph is characteristic of a robot path which was mostly piecewise-
linear; the robot goes straight for a while (the graph is horizontal), turns (the graph is
vertical), then goes straight again.
Gyros tend to drift (a roughly uniform linear deviation over time.) We try to calibrate
our gyros as carefully as we can so that the drift is small, but often our final data still
contains drift. Look at the horizontal segments of the graph: they aren’t quite horizontal.
Compute a better estimate of the gyro by finding a drift rate which makes the graph look
more like Fig. 9.
You can do this with code like this:

10
 4

−2

−4

−6

−8

−10

−12

−14

−16
0 1 2 3 4 5 6
4
x 10

Figure 9: Adjusted gyro data. The robot traveled in straight line segments, punctuated by
(roughly) 90 degree turns. This is readily evident from the gyro data.

>> gyro = unwrap(data(:,GYRO));

>> driftrate = <fill in this value>;
>> gyrocorrected = gyro + data(:,TIME)*driftrate;

Make sure you understand how this code works. Try it with and without the “unwrap”
command. Obviously, you’ll need to experiment with different values of driftrate. Dealing
with 2π wrap-around is tricky, and many bugs can be introduced by neglecting it!
A fancy way of doing the calibration is to find some metric that “measures” the goodness
of the drift rate. Suppose you assume that the starting orientation and end orientation were
different by π/2; could you compute the correct drift rate from this?

Problem 3. Basic odometry

Your next task is to replicate the odometry result in Fig. 8. This is the most “key” problem
in this whole tutorial, so don’t skip it!
Use the equations derived in this section to compute the change in θ at each time step.
You can integrate this by using the cumsum command. You’ll now have a vector of the
value of θ for every time step which you can use to compute the change in x and y. Using
cumsum again, compute the actual x and y position. Plot the result.
Your first attempt will look like Fig. 10. These results are not terribly good, but that’s
largely because the metersPerTick and baseline parameters were only estimates. Try to
improve the odometry result by adjusting the parameters. Remember that metersPerTick
might actually be different for the left and right wheel!
It may seem a bit tedious to optimize the parameters, and to a degree, it is. However, it is
possible (and desirable) to develop an intuition for how the parameters affect the odometry
result. You should be able to look at a result and say to yourself, “Hmm, I think the baseline
distance is too short.”
When optimizing parameters by hand, we often use the shape of the robot path as
our metric, mostly ignoring the scale. Can multiple sets of values for the baseline and
metersPerTick result in the same path shape?
You should also be aware of overfitting. The more “knobs” we have to adjust the size
and shape of our path, the more likely we are to be able to coax it into a shape we want–

11
 Uncorrected Odometry

10

5
Y (meters)

−5

−10
−25 −20 −15 −10 −5 0
X (meters)

Figure 10: Uncorrected Odometry. Your first attempt at odometry should match this figure.

regardless of what the data actually tells us. Our goal is to get the best general purpose
values of our parameters– ones that will be useful on future navigation missions, not get the
absolute best result on a single, controlled (and perhaps contrived!) experiment. Remember
that on any actual experimental run, it is much more likely that there was some error, than
that there was no error.
Even with only three parameters (leftMetersPerTick, rightMetersPerTick, and baseline)
it is quite possible to overfit, particularly when the dataset is relatively short. A general
strategy for dealing with overfitting is to use cross-validation: fit parameters to a subset of
the data, and test it on a disjoint subset of the data. We can’t readily do that here, since
we only have one relatively short dataset.

Problem 4. Gyro-assisted odometry

Suppose your robot is a tricycle, with only the center wheel having an optical encoder. You
no longer have dlef t and dright – you only have dcenter . This means that you cannot compute
changes in θ.
If you have a gyro, you could use the gyro’s estimate of heading instead! That is your
challenge in this problem. Using the gyro data that you adjusted in Problem 1, recompute
the odometry using only dcenter . Your graph should look similar to Fig. 11.
Is the path better or worse?
If you have dlef t , dright , and gyro data, how could you fuse the two independent estimates
of θ to get a better value of θ? If you’re interested in this, you should read about Kalman
filters. I recommend Chapter 1 of Peter Maybeck’s book, “Stochastic Models, Estimation,
and Control”; a PDF is available on the MASLab website.

Problem 5. Estimating velocity with back EMF

In the last problem, we assumed we didn’t have complete odometry data. Now we’re going
to assume we don’t have any!
The dataset contains a log of the PWM values given to the motor controller, as well as
the current consumption. A PWM value of -255 corresponds to about -12V, while +255
corresponds to +12V. The current consumption is given in amps, and the motor winding
has a resistance of 3.5 Ohms.

12
 Gyro−Corrected Odometry
18

16

14

12

Y (meters) 10

−10 −5 0 5 10 15
X (meters)

Figure 11: Gyro-derived heading odometry. The (drift-corrected) gyro was used exclusively
for heading. You should replicate these results, and try to improve them further by fusing
the odometry-derived heading and the gyro-derived heading.

Using Eqn. 17, estimate Vemf of the motor from these parameters. We know that Vemf
is proportional to angular velocity, but what is the scale factor? We could look it up on the
motor’s datasheet, but it’s probably not terribly accurate anyway.
Let’s consider a single wheel. We have it’s Vemf and its actual velocity as measured by
the quadrature phase encoders. To estimate velocity, you’ll need to consider the amount of
time that elapsed during every time step. You can compute this using:

>> dtime = [ 0; diff(data(:,TIME))]

One of the irritating things about working with real systems is that data doesn’t always
cooperate. It’s entirely possible (in fact, likely) that you may encounter intervals in which
zero time elapsed. When you compute the velocity during this interval, you must be careful
to avoid a divide-by-zero! A good way of dealing with this is to estimate the velocity by
looking at a window spanning several time samples. This can also serve to filter out some
of the noise!
We know that Vemf and velocity should be proportional to each other. Time for an
upsetting fact: there’s probably some nonlinear offset as well. In other words, let’s model
the system as:

velocity = x1 ∗ Vemf + x2 (18)

We want to solve for x1 and x2 , and we know the other two quantities.
This is (hopefully) true at every timestep, so let’s write this in matrix form. Let B be
the matrix of actual velocities, and let A be a matrix with Vemf in the first column and
1s in the second column (in matlabeese, this is A=[vemf ones(size(vemf))]). We can then
write:

Ax = B (19)
This of course is the most basic of all Linear Algebra problems. Remember that this
system is hugely overdetermined: we have thousands of simultaneous equations, and only
two unknown variables. This is a good opportunity to compute a least-squares fit. (For a
reminder of the mathematics, see Strang’s 18.06 textbook, “Introduction to Linear Alge-
bra”.)

13
 x = (AT A)−1 AT B (20)
We can now estimate the velocity of the motor using our coefficients x1 and x2 , and
compare it to our measured values using odometry. For the left wheel, you should get a
graph like that in Fig. 17.
4
x 10
2
Back−EMF estimated velocity
Quadphase velocity

1.5

0.5

−0.5

−1

−1.5
0 1 2 3 4 5 6
4
x 10

Figure 12: Estimating velocity from Current Sense. Using Eqn. 17, we can estimate back-
emf (and thus velocity) using the motor’s current and applied voltage.

Now let’s put it all together. Estimate the velocity for the left and right wheels, and
compute a trajectory without using the quadrature phase data (except for calibrating your
models.)
How does it look? When does it perform well/poorly? How does it look if you use the
gyro data for your θ estimate? You should be able to achieve something like Fig. 13. Keep
in mind that this approach doesn’t require any optical encoders at all!

Gyro Heading, EMF Velocity Navigation

18

16

14

12

10
Y (meters)

−2
−10 −5 0 5 10 15
X (meters)

Figure 13: Navigation from gyro and EMF-based velocity.

14
6 Conclusion
You now have a pretty good idea about what odometry is, what it can do, and what it can’t.
In Maslab, you won’t have optical encoders as good as those used in this problem. But you
will have a gyro, current sense, and the ability to build your own quadrature phase optical
encoders.
What approaches do you think will work? Is there some other way of navigating? Maslab
robots have cameras. If you pick out a feature and follow it around, can you use that to
compute changes in robot orientation? If you find a feature of known size, can you estimate
how far away you are?
You can purchase other sensors too– perhaps an electronic compass. Inexpensive elec-
tronic compasses might be low resolution, but they won’t drift like a gyro. Can an electronic
compass plus a gyro be a powerful navigation aid?
If you have any comments or corrections on this tutorial, please send me an email at
[email protected].

15