Cognitive Robot Architecture

ISAC’s cognitive robot architecture consists of several pieces of software agents within a multiagent-based software architecture called Intelligent Machine Architecture (IMA). An IMA agent is similar to an agent in multiagent systems in term of autonomy, communication, and pro-activity, according to characteristics of agents pointed out by [Franklin & Graesser, 1996]. However, instead of the more general approach of an agent representing a robot, an agent in ISAC represents only a piece of resource available to the system such as an actuator, or a computation function. These agents must work cooperatively in order for ISAC to perform a task. Designing the architecture in this manner alleviates difficulty in modification of functions in a complex system like ISAC.

Examples of key IMA agents are:

Perception Agents
  • Perception agents encapsulate hardware sensors and provide interpretation of the data to be used by other agents
Control Agents
  • Control agents take commands from other agents and send control signal to actuators
Compount Agents
  • Human Agent
    • Represents the status of the human interacting with the robot by interpreting his/her intention during a task execution. It also has capability to recognize human’s faces, and recall their preferences to use to interpret their intentions.
  • Self Agent
    • Represents the self of the robot through monitoring system status and agent communication. It also executes tasks using procedural and declarative knowledge, and controls behaviors using attention, emotion, and working memory.

In addition to IMA agents, ISAC’s cognitive architecture also consists of three types of memory structures including the Long-Term Memory (LTM), Short-Term Memory (STM), and Working Memory System (WMS).

STM
  • STM temporarily stores spatio-temporal data for sensory eventsand Implemented using Sensory Ego Sphere (SES) [Hambuchen, 2003]
LTM
  • LTM consists of three types of information: Procedural Memory (PM), Semantic Memory (SM), and Episodic Memory (EM)
WMS
  • WMS interfaces with to other memory structures and allows only task-related information to be stored. Therefore, it can be considered as a temporary “workbench” during a task execution
References

1. S. Franklin and A. Graesser, “Is it an agent, or just a program?” In J. P. Muller, M. Wooldridge, and N.R. Jennings, editors, Intelligent Agents III (LNAI Volume 1193), pp21–36, Springer-Verlag: Berlin, Germany, 1997.

2. Hambuchen, K.A., Multi-Modal Attention and Binding using a Sensory EgoSphere, Ph.D. Dissertation, Nashville, TN: Vanderbilt University, May 2004.