EXAMPLE PROBLEM

PRIOR

LIKELIHOOD

Remember that we are ignoring the denominator

Let us assume X1 and X2 are independent
Planning Uncertain Movements
•In robotics, uncertainty arises from partial
observability of the environment and from the
stochastic (or unmodeled) effects of the robot’s
actions.
•Errors can also arise from the use of approximation
algorithms such as particle filtering, which does not
provide the robot with an exact belief state even if the
stochastic nature of the environment is modeled
perfectly.
The field of robotics has adopted a range of techniques for
accommodating uncertainty.
• If the robot faces uncertainty only in its state transition,
but its state is fully observable, the problem is best
modeled as a Markov decision process (MDP). The
solution of an MDP is an optimal policy, which tells the
robot what to do in every possible state. In this way, it can
handle all sorts of motion errors.
•If the state is partially observable then the resulting robot
control problem is a partially observable MDP, or POMDP.
Unfortunately, techniques that solve POMDPs exactly are
inapplicable to robotics.
Robust methods
Uncertainty can be handled using so-called robust control methods
rather than probabilistic methods. A robust method is one that assumes
a bounded amount of uncertainty in each aspect of a problem, but does
not assign probabilities to values within the allowed interval. A robust
solution is one that works no matter what actual values occur, provided
they are within the assumed interval.

Fine-motion planning
Fine-motion planning (or FMP) is a robust method that involves
moving a robot arm in very close proximity to a static environment
object.
The solutions to FMP problems will typically be conditional plans or
policies that make use of sensor feedback during execution and are
guaranteed to work in all situations consistent with the assumed
uncertainty bounds.
 A fine-motion plan consists of a series of guarded motions.
Each guarded motion consists of
(1) a motion command and
(2) a termination condition, which is a predicate on the
robot’s sensor values, and returns true to indicate the end of
the guarded move.
The motion commands are typically compliant
motions that allow the effector to slide if the motion
command would cause collision with an obstacle.
Consider a two dimensional configuration space with
a narrow vertical hole. It could be the configuration
space for insertion of a rectangular peg into a hole or
a car key into the ignition.
The motion commands are constant velocities. The
termination conditions are contact with a surface. To
model uncertainty in control, we assume that instead
of moving in the commanded direction, the robot’s
actual motion lies in the cone Cv about it.
A two-dimensional environment, velocity uncertainty cone, and envelope of
possible robot motions.

The above figure shows what would happen if we commanded a

velocity straight down from the initial configuration. Because of the
uncertainty in velocity, the robot could move anywhere in the
conical envelope, possibly going into the hole, but more likely
landing to one side of it. Because the robot would not then know
which side of the hole it was on, it would not know which way to
move.
The intended velocity is v, but with uncertainty the actual velocity
could be anywhere in Cv, resulting in a final configuration
somewhere in the motion envelope, which means we wouldn’t
know if we hit the hole or not.

The first motion command and the resulting envelope of possible robot motions.

The second motion command and the envelope of possible

motions. Even with error, we will eventually get into the hole.
Because all possible velocities in the motion envelope are to the right,
the robot will slide to the right whenever it is in contact with a
horizontal surface.
It will slide down the right-hand vertical edge of the hole when it
touches it, because all possible velocities are down relative to a vertical
surface.
It will keep moving until it reaches the bottom of the hole, because
that is its termination condition. In spite of the control uncertainty, all
possible trajectories of the robot terminate in contact with the bottom
of the hole.
A fine-motion planner takes as input
•the configuration-space description,
•the angle of the velocity uncertainty cone, and
•a specification of what sensing is possible for termination (surface
contact in this case).
It should produce a multistep conditional plan or policy that is
guaranteed to succeed, if such a plan exists.
Dynamics and control
A rigid mobile robot has six degrees of freedom, three for its (x, y, z)
location in space and three for its angular orientation, known as
yaw, roll, and pitch. These six degrees define the kinematic state or
pose of the robot.
The dynamic state of a robot includes these six plus an additional six
dimensions for the rate of change of each kinematic dimension, that
is, their velocities.
A rigid mobile robot has six degrees of freedom, three for its (x, y, z)
location in space and three for its angular orientation, known as
yaw, roll, and pitch. These six degrees define the kinematic state or
pose of the robot.
The dynamic state of a robot includes these six plus an additional six
dimensions for the rate of change of each kinematic dimension, that
is, their velocities.
The dynamic state has higher dimension than the kinematic space would render
many motion planning algorithms inapplicable for robots. A common technique to
compensate for the limitations of kinematic plans is to use a separate mechanism, a
controller, for keeping the robot on track.
Consider an example where controller causes the robot to vibrate
rather violently. The vibration is the result of a natural inertia of the
robot arm: once driven back to its reference position the robot then
overshoots, which induces a symmetric error with opposite sign.
Such overshooting may continue along an entire trajectory, and the
resulting robot motion is far from desirable.

Robot arm control

Robot arm control Robot arm control using PD (proportional
using proportional control using proportional control derivative) control with gain
factors 0.3 for the proportional
with gain factor 1.0 with gain factor 0.1
component and 0.8 for the
differential component
 Controllers that provide force in negative proportion to the observed
error are known as P controllers. The letter ‘P’ stands for P proportional,
indicating that the actual control is proportional to the error of the
robot manipulator.
Let y(t) be the reference path, parameterized by time index t. The
control at generated by a P controller has the form:
at = KP (y(t) – xt)
Here xt is the state of the robot at time t and KP is a constant known
as the gain parameter of the controller and its value is called the gain
factor.
KP regulates how strongly the controller corrects for deviations
between the actual state xt and the desired one y(t).
In the above example,
KP = 1
Consider an example with
KP = 0.1
The simplest controller that achieves strict stability in our
domain is a PD controller. The letter ‘P’ stands again for
proportional, and ‘D’ stands for derivative. PD controllers
are described by the following equation:

As this equation suggests, PD controllers extend P

controllers by a differential component, which adds to the
value of at a term that is proportional to the first derivative
of the error y(t) – xt over time.
Robot arm control using PD (proportional derivative)
control with gain factors 0.3 for the proportional
component and 0.8 for the differential component
Clearly, the resulting path is much smoother, and does not
exhibit any obvious oscillations.
However, PD controllers may fail to regulate an error down
to zero, even in the absence of external perturbations. Wear
and tear in robot arms can cause systematic errors.

The solution to this problem lies in adding a third term to the control
law, based on the integrated error over time:
 Here Ki is yet another gain parameter. The term calculates
the integral of the error over time. The effect of this term is
that long-lasting deviations between the reference signal and
the actual state are corrected.
If, for example, xt is smaller than y(t) for a long period of
time, this integral will grow until the resulting control at
forces this error to shrink. Integral terms, then, ensure that a
controller does not exhibit systematic error, at the expense
of increased danger of oscillatory behaviour.

A controller with all three terms is called a PID controller

(Proportional integral derivative). PID controllers are widely
used in industry, for a variety of control problems.
Ethics and Risks of Artificial Intelligence in
Robotics
Many new technologies have had unintended negative side effects. If
the effects of AI technology are more likely to be negative than
positive, then it would be the moral responsibility of workers in the
field to redirect their research.
All scientists and engineers face ethical considerations of how they
should act, what projects should or should not be done, and how they
should be handled.
Some of the ethical risks of AI in Robotics are:
•People might lose their jobs to automation.
•People might have too much (or too little) leisure time.
•People might lose their sense of being unique.
•AI systems might be used toward undesirable ends.
•The use of AI systems might result in a loss of accountability.
•The success of AI might mean the end of the human race.
The above considerations lead inevitably to the
conclusion that we must weigh carefully, and soon,
the possible consequences of AI research.
The main aim should be to design a Friendly AI
Robot and the challenge is one of mechanism design
•to define a mechanism for evolving AI systems
under a system of checks and balances, and to give
the systems, utility functions that will remain
friendly in the face of such changes.
•Safeguards against change in utility function
should be built into AI systems.