COSPAL

The closest related project to the present considerations is COSPAL (Cognitive Systems using Perception-Action Learning) which can be considered a predecessor of DIPLECS. The COSPAL project aims at developing a new architecture for artificial cognitive systems which is based on perception-action learning to a largest possible extent. It attempts to reduce the innate or modelled capabilities to the absolutely necessary minimum. The demonstrator of the COSPAL project is fairly simple, as it consists of some cameras which observe the scene, a robot arm, and a shape sorter puzzle that is to be solved. This simple setting is static as the system itself is the only actively moving object, and it does not contain any kind of interaction, as the user controls the system by a GUI and not through the cameras.

Even though the COSPAL system is not designed for other tasks, the same implementation can be used for a simplistic control of a radio-controlled car. This observation leads directly to the objectives addressed in the DIPLECS project. The task of controlling a car (at whatever scale) principally extends the scenario of the shape sorter in three ways, that are not addressed in the COSPAL project

1. The scenario becomes dynamic, i.e., both the agent and objects in the environment
   move.
2. Interaction between different systems becomes necessary, i.e., other, actively controlled
   systems and humans exist in the same environment.
3. Real-time requirements for the control add another dimension to decision making: the
   system must send the next control command within a previously unknown time.

Extension 1 makes the low-level recognition process much harder to solve, as perceptionaction learning can no longer be based on the assumption that moving parts in the scene result from an action of the system itself. In COSPAL it is assumed that everything that is not stationary is part of or attached to the system’s manipulator. In the dynamic case, we have to associate a subset of moving objects to the actions, i.e., the simple object segmentation by instationarity no longer works. More advanced schemes have to be developed and we want to answer the question to which degree this can be based on learning only. The principles for incremental learning developed in COSPAL are crucial for the implementation of this mechanism.

Extension 2 makes the high-level part of the system much more demanding as the system has to analyse and anticipate the dynamic behaviour of other objects in the environment. The system must be able to abstract from recognising and analysing its own behaviour to recognising and analysing the effects of other agents’ behaviour. In terms of affordances and schemata, the system must map postulated affordances and schemata of other agents to itself. This includes also to communicate through actions with another agent, which introduces new challenges for the architecture design, but which might also lead to faster learning. Introducing the concept of external systems allows for learning from observation of other systems. Observing a system or human with a higher level of competence, leads to faster learning.

Extension 3 requires new strategies for decision making. A single-tracked feedforward-feed-back strategy for coming to a conclusion is no longer sufficient. Instead, the system must constantly provide some control signal, which potentially changes in each system loop, i.e., improves through time. The overall performance of the system no longer results from the architecture and learning time only, but also depends on the allocation of resources and finding good approximative reasoning schemes and nested strategies. The system must have the capability to evolve parallel, direct feed-forward reasoning branches, which take over after being trained sufficiently well compared to existing, more generic and slower paths.