Skip to this view's content

ETHx: AMRx Autonomous Mobile Robots @

edX

ETHx: AMRx Autonomous Mobile RobotsDashboard for: Vishay MORE OPTIONS

DROPDOWN Help
Courseware , current location
Course Info
Discussion
Wiki
Textbook
Progress
Course Overview
Introduction and Lecture Overview
Locomotion and Kinematics I (Legs)
Weekly Overview
Lecture Segments , current section
Problem Set
Problem Set due Mar 12, 2014 at 00:00 UTC
Locomotion and Kinematics II (Wheels)
[Optional] The AMRx Exercises
[Optional] Introduction to MATLAB
Previous
Next
INTRODUCTION

Skip to a navigable version of this video's transcript.

0:00 / 9:11
Fill browser
Turn off captions
Skip to end of transcript.
Caption will be displayed when you start playing the video.
DR. MARCO HUTTER: Hi.
Welcome back to Autonomous Mobile Robots.
My name is Marco Hutter, and I'm your teacher
for this segment on legged robotics, or in other words,
on machines that are built very similarly to humans or animals.
I'm going to start by giving you an overview about the different kinds
of machines that exist and elaborate on static and dynamic walking principles.
The goal is that you understand why are we building such machines,
how they work, and what the main challenges are to control them.
So let's start with an overview.
Considering the current landscape of legged robots
unveils a very broad spectrum of machines
with different skills and purpose.
This ranges from single legged hopping devices
that are used to study the fundamental principles of running to robots that
are built very similar to us humans or all kinds of animals.
The most obvious difference of all these machines
is the amount of legs and hence the number
of contact points with the environment they have,
which has a very strong influence on different aspects.

On one side, this influences the mechanical complexity

of the entire design.
On the other side, and this has often a large impact on the entire system,
the number of legs determines the way they locomote and hence
the control complexity.
This can be intuitively explained when comparing
the robots to their natural counterparts.
While animals like insects or spiders can walk directly upon birth,
mammals usually require several minutes to stand,
and we humans, the worst of all, need 7 to 10 months to crawl on arms and legs
and more than a year to stand and walk upright.
This example indicates that some gaits like crawling on four legs
are more stable and require less control action
while other gaits like walking on two legs
are more challenging in terms of control and coordination.
In robotics, we usually distinguish between static gaits and dynamic gaits.
Static gaits, on one hand, imply that the system
is stable at any point of time.
So whenever we stop the robot, it will not
fall but just remain in its configuration.
It is intuitively clear that the more legs a system has,
the better it is suited for static gaits.
On the other side of the spectrum, robots with a single leg
are incapable to perform static gaits.
In fact, single-legged machines like this early Raibert hopper
can only be stabilized on a so-called limit cycle.

And I want to illustrate this concept with a short movie in which we analyze
a planar, one-legged robot that is continuously hopping forward.
It's clear that, whenever the actuators are stopped,
the robot will just fall over.
However, by appropriate trusting during the ground contact phase and foot
placement during the swing phase, the robot
can maintain balance and propel itself forward
whereby every single jump is equal to the previous one.
The best way to illustrate this is by plotting
the state space of the main body-- so its vertical velocity,
horizontal velocity, and its height.
Now, to analyze the highly nonlinear dynamics,
we can focus on a single fixed point-- in this case,
the apex, which is the highest point in flight curve-and project the dynamics onto the corresponding transversal plane-the so-called Poincare section.
Now in this section, we can identify a first order
return map-- the so-called Poincare map-- which reduces a continuous time
system to a simple discrete mapping from one apex transit to the next.
We can see that the nominal gate always goes through the same point,
the so-called fixed point, which means that the robot returns always
to the same height and forward speed .
However, if we disturb the robot, the limit cycle
does not pass through this fixed point anymore.
And the question arises if it eventually comes back,
which means that the fixed point is stable,

or if it does completely diverge, which means that the system is unstable.

Let's recapitulate.
As we have seen, the complex, highly nonlinear dynamics
can be analyzed using the Poincare mapping from one apex to the next.
If we return to the same point, which means that xk equals to xk plus 1,
we call this a fixed point, x star.
To analyze whether this fixed point is stable or not,
we have to see how the errors at this fixed point evolve both from step
to step.
To this end, we linearize the system whereby the matrix dp
dx is called Monodromy matrix.
As we know from basic linear system theory,
we can analyze the stability of this fixed point by an eigenvalue analysis.
If all eigenvalues are smaller than 1, the fixed point is called stable.
As you've seen, single legged robots are very
useful to study basic principles of dynamic locomotion.
However, such robots have very limited applicability,
as they are incapable of standing still or performing static gaits,
and hence cannot work in rough terrain.
In my eyes, the best trade-off between static and dynamic locomotion
capabilities is achieved by a four-legged system.
One of the benefits is that these robots-- and this is in strong contrast
to bipedal- robots-- can be built with point feet
and do not need extended feet with actuated ankle joints.
As a result, we can keep the mechanic complexity low and the robustness high.
Furthermore, each leg requires only 3 degrees of freedom

to arbitrarily move the foot point.

Let's start with static walking on four legs.
In the gait, the body is always supported by at least three legs.
The center of gravity is kept within the support polygon.
And one leg is moved at a time to locomote forward.
With this gait pattern, walking looks as following.
The robot shifts its body to the new support polygon such
that it's standing statically stable.
Afterwards, the fourth leg is unloaded and brought
on a pre-defined swing trajectory to the new stance point.
As the name "static" already indicates, it's
the big advantage of applying such gait that the robot is stable at all time
and hence is well-suited, for example, to climb obstacles
as shown in this little-dog movie produced by USC.
However, such static gaits are very slow, and energetically often
inefficient, which is also the reason that you barely
see animals moving like this except in a very critical or challenging situation.
In most situations, animals apply dynamic gaits,
which means that there are phases in which the body is supported
by 2 or less feet at a time and that multiple feet are moved simultaneously.
As illustrated in this situation with the diagonal pair of legs
in simultaneous ground contact, the robot cannot produce torque around
the line of support, and hence will just fall over.
As a result, stability can only be achieved on a step to step basis.
So again, let's illustrate this with one of the robot
we developed in our group that is performing a trotting

gait over unpercieved obstacles on a treadmill.

As we can see, all gait characteristics are
kind of complimentary to the static gait.
Most importantly, and this is also something we can observe in this movie,
static instability is compensated by appropriately stepping to the side such
that the entire gait is stabilized.
I want to conclude this segment with a short overview, how
such system can be controlled.
A common approach is to first define the stepping sequence-- respectively,
timing of the gait.
For example, static walking, which means moving one
foot at the time-- dynamic trotting with the diagonal pairs of legs
and simultaneous ground contact-- or galloping.
In this case, a rotary gallop.
Second, the robot has to determine where to place its feet for two purpose-namely to maintain stability by propelling forward
and to react to terrain elevations.
This leg positioning is a pure kinematic control problem,
as we need to know how to adjust the motor position or velocity such
that the foot point follows a desired swing
trajectory to the goal stepping location.
Finally, the robot has to optimally distribute its contact forces
and joint torques such that it can balance and accelerate
the main body while not slipping and using as little energy as possible.
To know how to do this, we have to deal with system dynamics.
This short introduction to control concepts for legged robots

concludes this segment.

Thanks for your attention.
Go back to start of transcript.
Download video Download timed transcript
Download the slides of this lecture segment

Show Discussion New Post

Previous
Next
About Jobs Press FAQ Contact
edX LogoEdX is a non-profit created by founding partners Harvard and MIT whose mission is to
bring the best of higher education to students of all ages anywhere in the world, wherever there
is Internet access. EdX's free online MOOCs are interactive and subjects include computer
science, public health, and artificial intelligence.
edX on Meetup
edX on Facebook
edX on Twitter
edX on Google+
edX on YouTube
2014 edX, some rights reserved.

Terms of Service and Honor Code

Privacy Policy