1. VaMoRs, the experimental vehicle of UniBwM for 
autonomous mobility and machine vision, a 5-ton van. 
Always inexpensive PC-type computers have been used 
for the higher levels: initially, one PC based on the Intel 
80286 microprocessor in addition to the BVV 2 with 8086 
single board computers sufficed for guiding VaMoRs at 
its maximum speed of 96 km/h on an empty Autobahn in 
1987 exploiting the 4D approach. Only through the power- 
ful and intelligent interpretation constraints introduced by 
the integrated spatio-temporal models has it been possible 
to achieve these results with that low computing power on 
board. Since 1987 Intel 80386 single board computers 
have been installed on an intermediate hierarchical level 
in the BVV 2 [Mysliwetz, Dickmanns 87] resulting in 
much more robust road recognition under strongly per- 
turbed environmental conditions through shadows of 
trees. 
In 1991 all application software developed up to that 
point in different computer languages was translated into 
C and ported onto transputers. In a transition phase, both 
BVV 3 and transputers are used jointly; with the next 
generation of transputer processors the BVV will disap- 
pear. 
Since 1984 active viewing direction control has been 
applied in the framework of our vision systems [Mysliwetz 
84]. In 1986 it has been implemented for better recognition 
of curved roads [Mysliwetz, Dickmanns 86]. The micro- 
processor for viewing direction control is since integrated 
in the BVV 2. Especially with the introduction of a bifocal 
camera pair for better resolution further away this auto- 
matic viewing direction control became essential. 
2. The vision guided testbed ATHENE was built up in 
the year 1990 and is equipped with an almost identical 
sensor system as ' VaMoRs' except for the second TV- 
camera, but emphasis has been put on autonomous land- 
mark navigation. The operational environment has been 
provided with landmarks in the form of well discernable, 
static objects. Either the global position or the location of 
each target relative to the prescribed local trajectory has 
been known. The task of the real-time image processing 
system was to recognize the object and to deliver the 
corresponding measurement data to the navigation soft- 
ware. The event driven data fusion filter and a Kalman 
filter are used to combine different qualities of sensor data 
and to gain the best estimate of the robot's state. 
In case of ill conditioned optical information, the ve- 
hicle guidance system is able to travel a reasonable dis- 
tance between target sightings. This is a kind of 
’instrument flight’, realized with the memorized knowl- 
edge about the environment and the egomotion of the 
vehicle. 
The allowable distance travelled between optical up- 
dates is a function of how much drift from the nominal 
path is still safe for not colliding with an obstacle and for 
finding the next known landmark. 
Implementation for the AGV ATHENE started with the 
dead reckoning navigation approach. Reproduceable ex- 
    
  
  
  
  
  
  
  
  
  
  
  
   
  
   
   
  
  
   
    
   
  
   
   
   
   
    
    
  
   
    
  
   
  
  
    
  
   
  
   
   
  
   
  
  
    
  
   
  
   
  
  
    
  
N 
Fig.13: Demonstration experiment for landmark navigation 
periments showed, that it is possible on a smooth surface 
to travel over a distance of 20 meters with a lateral error 
of less than 1 cm. Another test course, shaped like an oval, 
showed that after a 16 meter ride and a 360 degree turn the 
heading error was less than 1 degree. 
These results have been obtained after putting some 
effort into the servo control mechanism. Stable and effec- 
tive control laws have been derived to allow accurate and 
safe operation of the vehicle. The control system is split 
up into different levels in order to have short reaction times 
for the vehicle to follow the trajectory commanded. But 
on rough ground, dead reckoning by itself does not yield 
any acceptable performance. 
After implementation of the landmark navigation mode, 
ATHENE moved autonomously around the laboratory 
area. The course consisted of four hallways with a total 
length of about 100 meters and a width of 1.80 meter. Four 
90 degree turns connect the hallways. The speed during 
an autonomous drive has been between 0.2 m/sec in 
narrow corners and 0.5 m/sec in straight hallways. Final 
experiments in late 1991 in a factory environment demon- 
strated the high precision navigation capability with visual 
feedback from landmarks. The task to be performed by the 
robot was the following (see fig.13): Starting position was 
at a roughly known location. The diameters of the error 
ellipses were between 10 and 25cm. After initialization 
with an artificial landmark (1) a straight line of work- 
benches on the right hand side had to be followed until 
reaching landmark (2); it consisted of a left turn corner. 
Next landmark (3) was an extremly narrow doorway ( 4 
cm free space at each side of the vehicle). A predifined 
path in a dead reckoning manner leads to the fourth 
landmark (4), which consisted of a closed door. A left turn 
brought the vehicle back to landmark (1), where it stopped. 
Then, a backward docking maneuver was performed to the 
starting position. The error ellipse now was less than 5 cm 
in diameter. The same course has been performed a second 
time after simulating a loading procedure.