A a 
another feedback law (e.g. preset decelaration rate) or by 
a prestored feedforward control or by a superposition of 
both. 
In a more refined version of this scheme, the under- 
standing of the situation may be differentiated to a more 
detailed level. If the road is wide enough and if the size 
and the position of the obstacle relative to the road can be 
estimated, the autonomous vehicle may be able to decide 
by itself whether the obstacle can be passed without leav- 
ing the road and touching the obstacle. This check and the 
corresponding special control activation can be performed 
in several different ways: 
a) The simplest one is to have some heuristic test 
procedure included in the code which works directly on 
image data; if certain patterns are matched the system 
could trigger some special control mode (parameterized 
feedforward) which guides the vehicle past the obstacle. 
This partially intelligent reaction is not very satisfactory 
in general; it may, however, be sufficient for certain 
applications. 
b) A more refined procedural approach determines the 
best estimate of the relative state of all objects involved. 
This combined state (the obstacle situation) is then ana- 
lysed using a preprogrammed classification scheme; as a 
result, feedforward or feedback or mixed control modes 
may be triggered for passing the obstacle. Special viewing 
direction control schemes may be invoked for careful 
feedback guidance of the vehicle past the obstacle. 
c) The last scheme will be treated separately in the next 
section since it involves explicit knowledge repre- 
sentation. 
The first two behavioral schemes can be subsumed 
under the blockdiagram of figure 11d. Depending on the 
number of behavioral rules implemented, relatively com- 
plex behaviors may be realized by this approach without 
resorting to explicit knowledge representation. It seems 
that in biological systems (animals) a similar scheme is 
widely used. Very well adapted motion behavior can be 
observed in rather nonintelligent species. 
Our autonomous vehicle 'VaMoRs', has demonstrated 
all its achievements using this rule based, switched direct 
feedback control strategy [Dickmanns, Christians 89; 
Zapp 88]. Convoy driving on a freeway and 'stop-and-go" 
in heavy traffic is the latest achievement using this scheme 
[Dickmanns, Mysliwetz 90]. Switching between feedback 
schemes with proper smooth transitions is the key to well 
adapted motion behavior. There is no direct interdepend- 
ence between the number of objects n, the number of 
available feedback control laws m and the number of 
feedforward control programs r. 
The approach developed is, from a functional point of 
view, similar to Brook’s subsumption architecture 
[Brooks 87]; however, all the subsumptions are realised in 
software based on a full spatio-temporal internal repre- 
sentation of relevant objects. This makes the system more 
flexible, allows easy changes of concepts and an evolu- 
  
  
  
Signal control 
Sensors EP transfor- Actuator |g 
| mation 
a) Output signal transformation directly into control actuation 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
State State feed- 
| Sensors estimation F1) back control Actuators === 
L 
b) Model based spatio-temporal state estimation (for one object) and 
reflex-like state feedback control (implicit notion of object state) 
  
  
  
  
  
  
Trigger — Event-trigge- 
pattern red feed 
forward 
. [7 
I control 
! 
State State feed- : > 
Sensors | estimation [7 v back control Actuators 
  
  
  
  
  
  
  
  
  
  
  
  
  
c) Event detection and triggering of a prerestored parameterized con- 
trol sequence for more flexible reactions (implicit notion of situation) 
      
    
  
  
     
  
    
      
  
    
   
situation analysis; 
control mode 
decision rule 
based 
  
  
feed-forward 
control program 
+ reflex 
ike 
tate behavior 
estimation 
  
   
       
feedback 
control law 
    
Sensors Actuators 
d) Selectable fast, reflex like feedback control determination with 
triggered feed forward components; situation dependent control mode 
  
situation assessment 
supervision gual orlented action planning knowledge 
adaptations (learniag) based 
  
  
  
   
  
  
  
  
   
      
— — D cm © - 
mode 
selection ul 
i Le e 
1 based 
feed forward ac 
progrema 
  
rule selection, 
moaitoring 
  
  
direct 
Deum reflex-like 
behavior 
  
faadbuek 
coutrol laws 
     
state 
estimallon 
   
e) Hierarchiecal scheme for adaptable fast control determination 
Fig.11: Steps in the evolution of intelligent control