Abstract: Ernst Dickmanns

Title: Dynamic Vision as a key element of AGI

A robust, flexible and extensible approach to intelligence for autonomous mobile systems is outlined, with dynamic vision at the center. The overall approach involves a synthesis of methods from cybernetics and AI, used to integrate closed-loop visual perception with prediction error feedback and goal-oriented action cycles using spatiotemporal models. In a layered architecture, systems dynamics methods with differential models prevail on the lower, data-intensive levels, but on higher levels, AI-type methods are used.
The first part of the talk reviews how in the 1980s these methods produced surprisingly high performance in visual guidance of autonomous vehicles with relatively low need of computing power. At the end of the EUREKA-project PROMETHEUS (1987-1994), several Mercedes SEL-500 vehicles demonstrated visually guided autonomous driving with guests on board on Autoroute 1 near Paris in typical 3-lane traffic at speeds up to 130 km/h, including convoy driving and autonomous lane changes. A1995 long-distance test drive involved more than 1600 km driven at speeds up to 180 km/h.
Based on these experiences, the third-generation dynamic vision system dubbed Expectation-based, Multi-focal, Saccadic (EMS-) vision was designed, and realized with standard Commercial-Off-The-Shelf PC-components. It is centered on the notion of subjects, defined as real-world objects with the capability to perceive and to act. Perceptual and behavioral capabilities (e.g. the own ones and those of other subjects observed) are represented in three interconnected parallel levels, thus allowing autonomous checks for improving robustness of interpretation and for system stability.
A number of applications of EMS vision beyond driverless cars are described and demonstrated in videos, including autonomous aircraft landing with checks whether the runway is free of obstacles, low altitude mission performance of helicopters, grasping of free-floating objects in Space, and assistance systems for highway cruising with mixed guidance, using both human (lateral) and autonomous control (longitudinal with radar and vision).
Next, the overall concept of dynamic vision based autonomous systems is reviewed, including key points with general applicability such as:

  • the deep integration of perception, decision making and actuation across three layers in the architecture via different methods in distributed computer systems;
  • the use of temporally extended adaptable maneuvers as mission elements, and
  • the modeling of individual subjects with specific capabilities, and objects with specific shapes and potential trajectories, to allow efficient adaptation to specific domains and situations.

The final portion of the talk turns to potential future applications, including autonomous mobile systems with increasingly powerful general intelligence. With the computing power available now, a variety of more complex applications are well within reach.