itrol
and
con-
ation)
avior
mode
wiedge
d
lex-like
avior
tionary path to higher developed decision and control
levels including knowledge based reasoning.
Explicit higher order world models
If the number of behavioral modes for different situa-
tions involving many different objects becomes larger and
larger it may become advantageous to structure the be-
havioral competences according to application areas and
classes of situations. The notion of goals to be achieved
becomes of importance during this process. This
developmental step may be the point where intelligence
proper comes into play for autonomous systems since it is
here that reasoning enters the field.
There is no more a direct link between a given situation
and the control mode selection for the lower level in
fig.11d. In the knowledge base there is now a set of goals
for the system determining the decision depending on the
situation. Some cost function to each goal yields a decision
criterion. Which control mode or which parameter set is
going to be applied depends on the actual minimal value
of several cost functions evaluated before decision taking;
those yielding the least cost usually will be the ones
selected. In order to avoid frequent switching in am-
bivalent situations, thresholding or temporal constraints
may be introduced more or less heuristically.
In fig.11e, parallel objects, modes and schemes of
fig.11d are shown for simplicity byrectangular boxes.
They encapsulate the basic cognition and behavior capa-
bilitiesof the system on which the highest knowledge
based level can build and which it exploits for realising its
plans.
By structuring the system in this way there is no need
for steady, especially fast reactions on the highest level
since well trained feedforward or feedback control appli-
cations on the lower levels are supposed to take care of the
continuous fast reaction components. The highest level
just has to do the monitoring and triggering.
Mission performance using landmarks
For the low speed AGV ATHENE the three-level
scheme of figure 11e has been slightly modified and may
3) slow navigation loop (event triggered)
2) trajectory guidance loop (100 ms)
| 1) fast inner control loop (< 20 ms)
: |
be shown as a cascaded triple feedback loop like displayed
in figure 12. In the outermost navigation loop the approxi-
mate direction of the movement is calculated from differ-
ent sources of a priori knowledge, but mainly utilizing the
job order (task map) and the environmental map informa-
tion (see fig.4 and 5). The job order tells the vehicle to
travel from a certain spot to a different location, mean-
while performing some given tasks. The environment map
(e.g. of the building) provides the heading direction, that
is the most convenient course towards the desired destina-
tion. With the help of the implemented simulation of the
whole setting it is possible to determine trajectories free
of collisions with known obstacles. Furthermore, only
those features for navigation will be marked in the land-
mark map, which will be visible during the real mission.
Depending on the operational mode, the reference tra-
jectory parameters are obtained either relative to an object
or as a predefined sub-task. Because of the positional
uncertainty the vehicle may have at the starting location,
the parameters for distance and direction will be corrected,
as soon as the real mission starts and the first landmarks
are in sight. All the long term planning is executed in a so
called mission planner module, which delivers a sequen-
tial list of single mission orders.
At the next level, a pilot-module will take these orders
and produce appropriate parameters for the path control-
ler. Vehicle path tracking is done by calculating a desired
heading angle, based on the mission order and the posi-
tional error. The pilot is responsible for navigation in the
local environment and performs its task together with the
state estimation module within a cycle time of 100 ms.
The control of the steering angle and the velocity of the
cart is performed by monitoring the signals of the gyro-
scope and the odometry. These specific control laws are
implemented on a seperate control computer; therefore,
the cycletime is less than 20 ms.
8. Experimental results
The approach described above has matured during half
a decade of experimentation with two experimental ve-
hicles at the university:
fast
feedback
controller
fast state
dynamics
of AGV
trajectory
dynamics
of AGV
d
,
S
mx
az
x Arajec 7 yj
Fig.12: Realization of visual landmark navigation
s ve System with multisensory feedback and high-level world ——
Maps, spatio-temporal models, behavioral competences
I
