Cognitive Architecture

AutoCrane has a central, predictive world model in which knowledge ("belief") about the system’s own state and of the environment is collected, filtered, integrated and updated (predicated) in a coherent form. A high deliberative layer makes fundamental decisions about the current goal to be pursued ("desire"), e.g. "Take waiting position", "Unload truck". In KAI, this layer is realized in the form of a finite automaton whose states represent the customer's process steps and whose transition edges represent simple predicate logic rules based on the information in the world model. Any commands from the outside (either from the crane operator or from the higher-level system control) are recorded as "external requests" and processed at this level.

KAI then breaks down the current goal into intermediate steps and then into concrete actions, down to the level of control commands to the individual axes of the crane. This part of decision-making and control is implemented on the basis of a hierarchical, behavior-based architecture. Individual behaviors ("open grapple", "enclose logs", "approach load") have preconditions, constraints and postconditions and are nested into complex strategies recursively using only two different arbitration methods: priority list, sequence. The arbitration methods select the individual behavior to be executed ("intent"), taking into account the current (intermediate) goal, the state in the world model and the pre- and boundary conditions. The selected behavior is then allowed to compute the control command to be sent to the crane.

The decision about the current (intermediate) goal is not made infrequently and then executed over longer periods of time as in classical planning approaches, but takes place in each individual cycle. In this way, the AI is able to respond immediately to changes in the system state or environment.

The central processing and command generation within KAI is organized sequentially and synchronously as in a classical control loop: the sensor information is processed, then integrated into the world model, and then a decision is generated from the current target to the individual behavior that is to be executed, which is then formulated into a concrete control command. KAI currently runs through this central cycle in the central software component with a cycle time of 50ms.

The world model contains a "Transformation Tree" - a method of kinematics (forward and inverse) borrowed from robotics [CRA04] - in which the relation of all joints and parts of the crane as well as the attached sensors is described and updated according to the movements. The Transformation Tree allows us to transform coordinates from the reference system of each sensor and joint to any other reference system. It is thus easy to transform measured joint angles into a position in Cartesian coordinates, or a target position in Cartesian coordinates into a position and movement in joint space.



Each command computed by KAI is checked for recognizably unsafe behavior before it is sent to the crane’s PLC. KAI uses a simplex architecture, common in safety engineering, where a safety overseer monitors the output, intervenes in case of violation of simple rules and thus modifies the system’s output in such a way that the safest possible state is achieved.