D2.1 Requirements and Specification - CORBYS

D2.1 Requirements and Specification
single element (logical or physical) of a hardware component, computational task or data. Software agents
communicate through asynchronous message passing. The agents are implemented as atomic or compound
agents. Atomic agents encapsulate a single resource and it does not depend on other agents. Compound
agents contain or depend on other agents for their primary function.
Sensing and Perception
In the IMA architecture each stimulus is processed by a different perception agent (different processing
techniques have been used for different sensors). To each sensor an IMA agent that processes sensor
information and stores the result in short-term memory is assigned. For example, there are separate visual
agents that perform recognition of objects, object localisation, face recognition and etc. The modular
approach makes it possible to add additional sensors.
Human-robot Interaction:
The two software agents are responsible for HRI (Kawamura et al., 2000), one of them is human-agent that
encapsulates information that robot determined about human and other is self-agent that addresses the
humanoid’s cognitive aspects.
Real-time control:
The initial usage of IMA architecture in the ISAC caused problems with real-time control (Kawamura et. al.
2004). Asynchronous message passing has caused latency in passing control inputs to real-time controllers.
Therefore, separate encapsulation of multi-agent tasks (such as visual serving) has been carried out. Head,
arm and hand agents are responsible for controlling the head, arm and hand actuators. They accept commands
from one or more clients and carry out command arbitration. The agents also provide the clients with
information about its current state (for example joint position). The actuation agents pass referent joint angles
and velocities to servo controllers. The servo control loops have been realised using a QNX real-time
operating system.
Planning:
The self-agent is responsible for planning actions of the robot. The most important modules of the Self-agent
are Central-Executive-Agent (CEA) and First-Order-Response (FOR). CEA module makes decisions and
invokes skills necessary for performing the given task that has been selected by Intention-Agent based on
perceived sensory information. The FOR module is responsible for creating the reactive responses of the
robot.
Learning:
The system is equipped with different types of memories in order to support learning. The memory structure is
modular and divided into three components: short-term memory, long-term memory and working memory.
Short-term memory holds most recent sensory information on the current environment in which ISAC robot
operates. Long-term memory holds learned behaviours, semantic knowledge and past experience and it has
three structures: semantic, procedural and episodic. Working-memory holds task-specific information that
authors call “chunks”. The working-memory uses temporal difference learning algorithms and neural network
to provide learning in IMA architecture.
Communication:
The communication between agents is based on asynchronous message passing. Any agent can be accessed
by any other agent. By increasing the number of agents the system faced with communication “lock-ups”
13.2.3 iCub
The cognitive architecture of iCub robot is depicted in Figure 33. The architecture “..comprises a network of
139