Robotics

• Biomorphic robotics and otherwise
• Basic approaches
  • Classical architectures: central control; manipulation of explicit knowledge; perception (sensing), reasoning and planning, and action modules (decomposition by function)
  • Situated, behavior-based architectures (Brooks): eliminate explicit deliberation; special-purpose, task-oriented modules (decomposition by activity)
• Robotics involves planning (one of those topics we don't have time for), but it is harder than just planning because the real world is
  • Inaccessible: you can't know everything you need to know.
  • Nondeterministic: the state of the robot itself, as well as of observable aspects of the world, is not completely predictable.
  • Dynamic: (1) what worked once may not work later on (2) the world may change while you're planning, rendering the plan useless.
  • Continuous: you can't enumerate all the possible states or actions.
• Robotics is not just action; it also requires sensing: proprioception, force sensing, tactile sensing, sonar, vision
• Kinematics: correspondence between geometry of actuator and geometry of world
  • Effectors and degrees of freedom
    • Effectors for locomotion and manipulation
    • DOF for object in space (6), simple object on a flat surface (3), car with trailer (4)
    • DOF for car (2); turntable (3); joints: hinge (1), prismatic (1), rotary (1), ball-and-socket (3), human arm (3+1+3=7)
    • When the number of controllable degrees of freedom is less than the total number, it is relatively difficult or impossible to reach certain states.
    • When the number of controllable degrees of freedom is the same as the number required to specify the end state, there is one set of settings for each end state
    • When the number of controllable degrees of freedom is greater than that required to specify the end state, there are multiple ways to reach end states
    • Problem of interaction of degrees of freedom of actuator
  • Forward kinematics: given a set of joint angles (or some other set of settings of the system), specify a resultant configuration
  • Inverse kinematics: given a goal configuration, specify a set of joint angles (or other settings); problem of indeterminacy when number of degrees of freedom in system exceeds number of degrees of freedom needed to specify task
  • Examples of inverse kinematics problems
    • Getting an arm into a desired position and orientation
    • Producing a given linguistic sound with a set of articulators
  • Approaches to inverse kinematics

    
    (y* is the goal state, y the state which is actually reached; with no feedback x = x')

    • Use feedback: move the system in the desired direction, correcting errors over time (often too slow for humans)
    • Reinforcement learning: given a particular input, increase or reduce the probability of a particular output in response to reinforcement signal
    • Direct inverse modeling: observe the I/O behavior of the plant, then train on the inverse for the feedforward controller
    • Forward modeling: first learn a model of the plant, then learn the parameters of the feedforward controller