Maurer and Dickmanns 1997, Abstract: An advanced control architecture for autonomous vehicles is presented. The hierarchical architecture consists of four levels: a vehicle level, a control level, a rule-based level and a knowledge-based level. A special focus is on forms of internal representation, which have to be chosen adequately for each level. The control scheme is applied to ‘VaMP, a Mercedes passenger car which autonomously performs missions on German freeways. VaMP perceives the environment with its sense of vision and conventional sensors. It controls its actuators for locomotion and attention focusing. Modules for perception, cognition and action are discussed.

1. INTRODUCTION
The problem of road vehicle guidance by computer vision has been addressed in the US and in Germany since the early eighties; in 1986, first test trials with the 5-ton van VaMoRs were made on a test track. One year later, VaMoRs was able to perform road running at speeds up to 96 km/h over more than 20 km distance on a freeway not yet opened for public traffic. In 1994, the second generation of autonomous land vehicles developed in close cooperation between UBM and Daimler-Benz was presented to the public. The passenger cars (Mercedes 500 SEL) VaMP (UBM) and VITA2 (Daimler-Benz) demonstrated autonomous driving on a public freeway. In terms of robustness, the peak performance in autonomous road vehicle guidance on freeways was demonstrated on a long distance drive from Munich to Odense in Denmark. During the test drive 1530 km, a percentage of 94.4, were driven with automatic lane keeping and automatic longitudinal control. More than four hundred lane changes were performed by the vehicle itself after being triggered by the human operator.
   From the very beginning at UBM the development of perception algorithms and behavior has been discussed in the context of a mission. Wuensche presented his work on dynamic machine vision in the context of a mission: An air cushion vehicle had to dock on several objects in a technical environment. Hock developed a navigation module which controlled missions of unmanned transport systems and road vehicles in a static environment. In cooperation with the computer science department of UBM the feasibility of autonomous decisions for lane changes was demonstrated on a public freeway. In this application the ‘situation’ has become a central concept.
   In this paper a system architecture is presented which summarizes the experience accumulated at our department. The focus of the paper is on the following items: The rough structure of the system architecture should be independent of the task and, as far as possible, independent of the vehicle.
The architecture consists of four levels, which differ in the abstraction of representation: a vehicle level, a control level, a rule-based level, and a knowledge-based level.
   A synthesis of methods from control theory and computer science is given. Explicit representation of the capabilities in both perception and action is regarded as crucial for the overall performance of the autonomous system. Decreasing the degree of autonomy is proposed as an appropriate reaction when problems in the perception modules occur (graceful degradation of the system’).