Previous: The Sensori-Actuator level. Up: The agent Next: The Knowledge Level.

The Perceptuo-Motor Level.

At the Perceptuo-Motor level, primitive sensations and actions are processed and combined to form more complex perceptions and behaviors. We list some of these below. We are currently investigating the use of foveal vision for the agent, which would alter some of the descriptions below.

A signal, object-seen, is generated to denote having seen an object. Signals will be generated to denote that the object is in the right, center, or left field of view. If an object in the field of view becomes larger (due to getting closer), a signal is generated that the object is object-bigger. Similarly, the object in the field of view will get smaller when departing from the object, signaled by object-smaller. If the object becomes an obstacle for the robot, the vision system generates an object-too-close signal. If anything moves in the robot's peripheral visual field, it will generate the signal moving-object (together with a position signal). Another complex perception is the trajectory of robot movement. As the robot senses its wheel speed and direction of rotation combined with visual input, it builds a trajectory of movement. This can be used in building a map of the domain.

When the robot is in motion, it will use cues from its environment to guide its behavior selections and subsequent learning. In order for the robot to guide its behavior, we need to associate rewards with actions that result in desirable sensations. For instance, if the robot wants (at the Knowledge level) to touch an object and is taking actions to move towards the object, and the object is in the left field of view and the robot moves left (increases its right wheel motor speed) it will bring the object to the center of the field of view. The action of turning left in this situation will be positively rewarded. The result of learning (sequences of) actions will be recorded as a PMA. For instance, touching an object will evolve into a PMA. Even after the initial learning of a PMA for a complex action, learning will continue to improve the PMA.

For each behavior, a triple of <A, S, R> will be defined. A is the set of primitive actions, S is a set of sensations, and R is a set of rewards. For example, for the behavior of touching, A = {left wheel forward increase speed, left wheel forward decrease speed, right wheel forward increase speed, right wheel forward decrease speed}; S = {object is bigger, object is smaller, object is in the left field of view, object is in the right field of view, object is in the center of the field of view, object is too close, contact is made}; R = {object is bigger , object is smaller , object is in the left field of view , object is in the right field of view , object is in the center of field of view , object is too close , contact is made } (numbers ranging from to are rewards with denoting desirable and denoting undesirable).

Below is a list of behaviors and percepts we want the robot to learn. We will provide the robot with appropriate rewards for these behaviors. Below we give a list of emergent behaviors that the robot will learn completely on its own (no rewards).

• Touch behavior: approach an object until contact is made.
• Proximity percepts: proximity to an object, derived from camera data. (e.g., see all the signals we listed above in the example of touching behavior).
• Approach behavior: approach an object until it is in proximity.
• Block percept: recognizing something in the field of view as a block.
• Push a block
• Find a block
• Unwedge a block
• Explore/map the room

When more than one agent is present, behaviors at this level may include finding or hiding from other agents, following or running from other agents, playing games (hide and seek, for instance).

Earlier we defined behaviors in terms of PMAs. We consider subsets of PMAs to also be behaviors. Percepts and behaviors listed below are emergent, i.e., they are learned but the agent did not intend to learn them. The robot is never told about these behaviors. They are learned in PMAs, as in the PMA for touching objects.

• Proprioceptive movement percepts and movement behaviors: moving forward, backward, turning left forward, turning right forward, turning left backward, turning right backward, rotating left, and rotating right. Each of these is a percept as well as a behavior.

Some other percepts and behaviors situated at the Perceptuo-Motor level are listed below. They are hard-wired in the initial implementation, but could conceivably be learned or emergent as well. As a rule of thumb, we consider the perceptual representation required to perform discrimination tasks to be hard-wired, and those required to perform identification tasks to be learned (or learnable).

• Color percepts: in perceptual color space coordinates, derived from camera data.

• Color hallucination behavior: imagine the surroundings in different colors by warping the perceptual color space.

• Shape percepts: simple object shapes like square, rectangle, tall, short, narrow, wide, irregular, big, small, derived from camera data.

Also at this level, basic emotions like fear and curiosity may be implemented as particular types of behaviors (aligned with symbolic labels at the Knowledge level).