TeleoR Robotic agents: Deductive Belief Store, robotic resource actions, sensor percepts, inter-agent communication

A TeleoR robotic agent is a multi-threaded software process that controls of one or more robotic resources, e.g. one mobile robot, two robotic arms, possibly alternating their use between several concurrent tasks. It controls the robotic resources using a repertoire of control actions, such as turn(left,0.4), pickup(arm1, block2), in order to perform its tasks. The devices exist in a real or simulated environment external to the agent. The tasks have goals to bring about, or to maintan, or to prevent, certain states of the agent's environment. The agent does this by frequently sensing the environment of the robotic resources, and by responding with appropriate resource actions, as the sense data changes. Other agents may be in the same environment controlling different robotic resources. They could be working on completely independent tasks, or co-operating on some or all the tasks.

The sensors may be attached to the robotic resources, or be completely independent. A camera mounted on a mobile robot, or at the end of a moveable telescopic arm, is an attached sensor. A wall mounted temperature sensor, or a camera that automatically pans using a fixed trajectory, sending video images at frequent but fixed intervals, is an independent sensor. The readings of the sensors are converted into percepts: facts such as close(bottle,left) and on(block2,block4). This conversion is done by software external to the TeleoR agent. Each percept is a logical interpretation of one or more sensor readings that is relevant to the agent's tasks. The facts are handled by a special percept handling thread within the agent. At a frequency required by its tasks, this thread atomically updates the agent's Belief Store BS. During the update no other agent thread is active. The BS facts are what the agent believes about the current (and perhaps past) state of its environment, about other robotic agents controlling robotic resources in the same environment, and the robotic resources that it and the other agents control. Non-percept belief facts may have been distilled from messages it has received from other agents, or may be facts encoding memories of past perceptions and actions.

Reasoning agents also have a separate knowledge layer KL of fixed knowledge rules and facts. The KL facts encode fixed properties of the environment and other agents relevant to the agent's tasks , e.g. that a red door connects two named rooms. They are fixed in that they will not change during the agent's lifetime. The KL rules are used to draw inferences from BS and the KL facts. They encode information about the semantics of the predicates used in these facts and define new task relevant fusions of the facts. Facts about the environment that might change as a result of exogenous events, such as which doors are open or closed or blocked, are dynamic beliefs. The dynamic beliefs and the knowledge rules and facts comprise the agent's deductive belief store, which we shall henceforth call its DBS.

Often all the robotic resources and sensors are on one robot, e.g. a small mobile robot with a forward facing camera, bump sensors and a gripper with touch sensors, and the agent is a process on a computer inside the robot. Alternatively, the robotic resources and sensors could be distributed around the rooms of a house with the agent being a process on a laptop using WIFI.