Attention Subsystem

From TinyCog
Jump to: navigation, search

The attention subsystems controls the "focus of attention" of the Scene Based Reasoning system. The focus of attention corresponds most of the time with the "camera position" of a robot running SBR. However, the attention focus can also be directed to senso-motoric constituents of a system including memories from the Episodic Memory, the execution of plans and other internal structures.

The sensor input acquired from the area focused on will be passed through 3D Reconstruction and directed to the Episodic Memory in order to retrieve "associations", i.e. plans and scripts associated with the focused objects and their context. The attention subsystem retrieves these "ideas popping up" and tries to match them with the current goal hierarchy in order to determine if the idea could contribute to solve any of the currently active plans. Matching the idea against the active plans deviates the attention away from the current focus and towards internal objects.

Attention Focus

The selection of a scene or script as a focus for the SBR system's processing capacity. Most of the time the attention focus lies with the images from a camera of a robot running SBR, but attention can also be focused on scene representa-tions of plans and other SBR structures allowing for various levels of introspection. The focused sensor data are processed by 3D reconstruction and passed on to the episodic memory in order to retrieve "associations", i.e. plans and scripts associated with the focused objects and their context. The attention subsystem retrieves these "ideas popping up" and matches them with the active "persistent plans" in order to determine if the idea could contribute any of the currently active plans. When executing plans or "active tasks", the attention focus tracks the current vs. planned world state and initiates re-planning if necessary.

Comparison and References

[Langley 2005] proposes a "Meander" subsystem for the ICARUS cognitive architecture with similar execution tracking and re-planning properties.

Implementation Status

TinyCog 0.0.1 does not yet implement the attention subsystem. Implementing the Hunter Domain will require the tracking of medium-term plan, which is only a part of the attention subsystem.