Perception in BBR

• Classical AI: perception independent of action
• remember: idea was to construct a model of the world
• BBR: perception dependent on action
• idea: perception is in the service of action, this means:
• tuning perception to the needs of the motor activities
• use knowledge to constrain perception (what is out there and where it is)
• use motors to enhance perception (active perception)
• divide into different perceptual classes

• In general: BBR advocates a view of perception, where only the sensory information necessary for action needs to be processed (not all the available information--e.g., detect all objects in a scene)
• distinguish:
• exteroceptive vs. proprioceptive sensors
• ideothetic vs. allothetic sensory information (path integration)
• open-loop vs. closed loop control
• efficient processing sensory information in parallel (by avoiding the integration of information into one "world model")
• Also: symbol grounding can be solved

• The BBR view is not new:
• ecological psychology (J.J.Gibson)
• "Affordances"--refers to the action an object "affords" (i.e., what can be done with it)
• Crucial claim: perceptual systems "pick up" affordances directly
• no additional processing or reasoning required!
• that's why this view is attractive to BBR (often called "perceptual schema")
• define perceptual classes for particular needs
• action-oriented perception (Neisser)
• at the heart: perceive-act cycle (which modifies behavioral propensities and dispositions as well as expectations)

Sensors

• Distinguish:
• active (emit energy into the environment)
• sonar
• laser
• passive (use energy available in the environment)
• "dead reckoning"
• shaft encoders--count number of rotations of motor shaft
• INS (inertial navigational system)--measures acceleration
• vision (CCD cameras)

Combining Action and Perception

• Remember: in BBR action and perception are viewed as intrinsically intertwined
• Hence, perception and action can be used to mutually constrain each other
• typical in AI: use perception to limit/improve action
• especially in BBR: use action to limit/improve perception
• action-oriented perception: behavioral needs determine the applied perceptual strategies (Neisser, Arbib)
• process only that part of the sensory information that is relevant and required for the current behavior
• distinguish:
• sensor fission (what are advantages/disadvantages?)
• sensor fusion (what are advantages/disadvantages?)
• sensor fashion (what are advantages/disadvantages?)
• active perception: perceptual requirements dictate the robot's actions
• perform actions that support and can possibly enhance perceptual activity
• sample case: active vision (use motor capacity to explore the visual field)

• The turn in BBR: maybe models are not so bad after all? (e.g., Brooks)
• In particular, use knowledge about the world to constrain perception (what is out there and where it is)
• expectation: "what to look for" 
  <fill in>
• attention: "where to look for it"
  <fill in>
• sequencing: "when to look for it"
  <fill in>
• configuration: "how to look for it"
  <fill in>

• Important questions for the design of perception/action subsystems:
• what perceptual information is necessary for task?
• what information is sufficient to achieve task with given action control?
• what apriori knowledge can constrain perceptual tasks?
• what are the sensory limits (in principle and in practice)?
• how can action improve perception in a given system?