• PART 1 : OVERVIEW of ROBOTIC PARADIGMS
• CHAPTER 2 : THE HIERARCHICAL PARADIGM
• CHAPTER 3 : BIOLOGICAL FOUNDATIONS OF THE REACTIVE PARADIGM
• CHAPTER 4 : THE REACTIVE PARADIGM
• CHAPTER 5 : DESIGNING A REACTIVE IMPLEMENTATION
• CHAPTER 7 : THE HYBRID DELIBERATIVE/REACTIVE PARADIGM

• First robot term: in Prague, 1921, a biological creature in the play, with no right, only serves to human. Czech term, 'robota' which means menial labor.
  
• Main science fiction films:
• Metropolis (1926)
• The Day the Earth Stood Still (1951)
• Forbidden Planet (1956)
• After used in factory's : mindless and good only for well defined repetitous types of work.
• The definition : an intelligent robot is a mechanical creature which can function autonomously.
• Function Autonomous : adapt changes in its environment, without requiring human operation.

• A paradigm is a philosophy or a set of assumptions and/or techniques which characterize an approach to a class of problems. Robotic paradigms are defined by the relactionship among 3 basic primitives : sence, plan, and act. Other was to define the paradigm is to look how sensory data is distributed and processed.
  
• Hierarchial Paradigm :
  • oldest paradigm
  • heavy planning, a top-down paradigm
  • in each step, robot make plans
  • robot construct a global world model, which has a closed-world assumption and suffers from frame problem.
• Reactive Paradigm:
• Popular mainly from 1988-92, since then hybrid model was used more.
• Two main trends among researchers :
• They investigate biological and conginitive sciences for living creature behaviors,
• Decrease in cost, fast execution time.
• Behaviors are composed of sense->act couplings and behaviors combine for the actions.
• Throwing away planning phase was not good for general purpose robots.
• Hybrid Deliberative/Reactive Paradigm:
• Task is divided into subtasks, behaviors are determined. plan-> sense-act

• First AI robot was Shekey, 1967, which was built at SRI, DARPA.
• Eyes open, sense and construct the world. Eyes closed and plan the directives. Acts to carry out first directive. In a cycle...
  
• World model composed of, a priori knowledge (map of building), sensory information and any additional cognitive knowledge for accomplishing the task.

• Strips is a general problem solver. It uses means-ends anaysis. If it cannot find the solution of the problem in one step, its selects an action to reduce the difference of states, tries to approach to final state.
• There is an initial state, final state, difference between states. Difference evaulator takes the differences. Different operators are selected (fly, drive,walk). A difference table is drawn which has difference and operator columns. Robot operates according to the current difference value.
  
• But in the middle of the program, there should be some pre-conditions added to the table, for example drive_rental if robot is at airport and drive_personal if robo is at home.
  
• But how could robot know, the location of self? Then in each step there should be information that should be inserted and deleted to/from the robot knowledge. This is performed by inserting columns to the table.
• A program can be written to step by step accomplish the operators to reach the final state.
  
• Predicate logic is used in Strips. For example INROOM(IT,R1) is a predicate which can be inserted into knowledge base (world model), if robot is in the room 1. In order to go through the door from room 1 to room 2, there should be some preconditions in the world model, one of which is INROOM(IT,R1). After this action, INROOM predicate should be deleted from knowledge base, using the difference table. Since Strips tries to reduce the differences between initial and goal state, it should look for the add-list column of difference table, while selecting the operator. In this step, difference operator is used.Negating the difference...
  

• An architecture is a method o implementing a paradigm, of embodying the principles in some concrete way.
• Nested Hierarchical Controller: Mission planner, Navigator and Pilot submodules are hierarchically located in planning module. Robot moves eyes open wide, for any unexpected event, it is not needed to rerun all planning module. The disadvantage is, it is only appropriate for navigational tasks.

• For specific applications.
• Simple for the ordering relationship of SENSE-PLAN-ACT.
• Primary disadvantage was planning.
• Slow planning process : a falling rock on the robot...
• Strips should reason all irrelevant details of the world model. NHC and RCS tried to divide world model, but they were so specific.
• Uncertainty is a very important issue for hierarchical paradigm.

• Micheal Arbib was one of the first who study on animal intelligence models.
  
• Valentino Braitenberg : Vehicles.
• Concurrent simple animal behaviors interact to produce complex behaviors : Lorenz and Tinberger.
• Foundation for thinking about robotic perception from cognitive way : J.J. Gibson and Ulrich Neisser. - affordance
• Schema theory was applied successfully by Arbib to represent both animal and robot behavior.

• Simple animals live in an open-world, they have no brain, but they can overcome the frame-problem.

Agency and computational theory

• Agent: self-contained and independent. Interact with world to make changes and sense what is happening. Self-awareness.

• Computational theory defines the commonality levels of intelligent entities.

  • Sharing a commonality of purpose and functionality in the most abstract level.

  • Decomposition of input, output and transformation modules. Black box of transformation is unknown and not considered. Common agents exhibit common processes.

  • Implementation of the process thru black box. Agents (an animal and a robot) may not have any commonality. Duplication is not wise.

What are animal behaviors?

• Ethologists work on animal behavior. Behavior is a mapping from sensory input to motor actuators. Behaviors divided into 3 broad categories:

  • Reflexive Behavior: Its mapping in robotics is reactive paradigm. Hardwired, very fast, no processing is needed.

    • Reflexes : Response lasts only as long as stimulus and response is promotional to the intensity of stimulus.

    • Taxes : Response is to move to a particular direction according to stimulus orientation. (baby turtle-tropotaxis)

    • Fixed Action Patterns : Response lasts for a longer time than the stimulus.

Coordination and Control of Behaviors

• Konrad Lorenz and Niko Tinberger were the fathers of ethology. They work on set of coordinated behaviors.

  • Innate : Born with a behavior. Arctic terns, if baby tern is hungry -> peck at the largest red blob in the sight. no identification of the parents.

  • Secuence of Innates : Born wirh sequence of behaviors. Female wasp mates with male -> build nest. When nest is completed -> lay eggs. Each step is triggered by the combination of internal state and the environment. (stimulus releasing)

  • Innate with memory : Baby ants are curiously born that, step by step they try to learn the outside view of the hive. The behavior of zooming around the hive is innate; what is learned about the appearance of the hive and where the opening is requires memory.

  • Learn : Learn how to hunt of baby lions. Hunting is a complex process.

• The studies made, internal state and motivation of the agent is important while releasing a behavior. Other lesson is behaviors can be sequenced to create complex behaviors.

• Innate Releasing Mechanisms : An IRM activates the behavior. A releaser is a boolean variable to be set. It acts as a control signal to activate a behavior. Programming the execution order of behaviors due to the releasers is a difficult task, using if-else structure for example. There should be releasers, compound releasers, implicit chaining, precedence and behavior blocks, time insertions (problem of out of view predator when turn the predator back for fleeing.). Mainly inhibition, T-periods, and interrupts make program less general purpose. So independent simple behaviors needed.
  

• Concurrent Behaviors : Some behaviors may execute concurrently and independently. Some behaviors may violate or ignore the implicit sequence when the environment presents conflicting stimuli.

  • Equilibrium : Balance each other. (squirrel cannot decide to flee or feed)

  • Dominance : Winner take all. (hungry and sleepy)

  • Cancellation : Cancel each other.

Perception in Behaviors

• Action perception cycles : Cognitive action ->[directs what to look for]-> Perception of Environment ->[agent samples]-> World-> [agent modifies world] -> Cognitive action..
  

• Two functions of perception

1. releasing behavior
2. guiding behavior
   

• Gibson : Ecological approach ( affordance ) : "The world is its best own representation." Affordances are perceivable potentialities of the environment for an action. Baby arctic tern, color red is perceivable and represents potential for feeding. Affordance is only potential, counts for other conditions.

  • Poppy color is evolved for certain bandwidth and bee retina is evolved to sense that color.

  • Affordance is directly perceivable.

  • Affordance and specialized detectors are very difficult to duplicate.

  • Chair discrimination problem, instead of using a structural approach to the problem, functionality measure is used. Affordance of sittability. Stark and Bowyer made very successful computer program which use affordance.

• Neisser : Two perceptual systems :  a) direct perception accounts for affordances which evolved earlier. b) recognition is used when we are discriminating 'my coffee cup' from 'others coffee cup'. This is top-down, problem solving and cognitive.

Schema Theory

• Schema theory is the means of casting above discussions and insights into object oriented programming.

• First used by Michael Arbib.

• Used by Arbib and Murphy for mobile robot control, Lyons and Iberall for manipulation, and Drapel for vision.

• Schema represent the basic unit for activity. It is a generic class which contains data (knowledge, models, releasers) and methods (algorithms of perceiving and acting).

• Riding bicycle is a template but you can ride different bicycles. A creation of a specific schema is called schema instantiation. move_to_food is a schema which has a template activity of 'always head in a straight line', but a candy or sandwich may be different instantiations.

• Behaviors and schema theory:

  • A behavior is a schema which is composed of at least one perceptual schema and one motor schema.

  • Releasers should be inserted into data part of a schema, if RMS's should be used, with a IRM_logic unit.

  • Schema theory is expressive enough to represent basic concepts like IRM's, plus it supports building new behaviors out of primitive components.

  • Different perceptual schemas may work independently according to the environment state.

  • For a toad's feeding behavior, one can construct a behavior using schema theory as: Releaser: appearance of moving object. Perceptual Schema: get coordinates of the moving object. Motor Schema: turn to coordinates of small object.

  • Since output is represented as a vector field, when two flies are in the perception field, two distinct feeding behaviors are triggered and vector's are summed up.
    

  • Lesson : Functionality in an animal and a computer can be equivalent.

  • Inhibition mechanisms or combination of different behaviors are represented by combination of different schemas.

Principles and Issues in Transferring Insights of Robots

• Decomposition of complex actions into independent behaviors.

• To simplify control, agent should rely on boolean activation mechanisms.

• For the computational expense of sensing, perception must filter that only relevant information should be taken. (action-oriented perception)

• Direct perception reduces the computational complexity of sensing.

• Behaviors are independent but may be combined or inhibited.

• The resolution of conflicts between concurrent behaviors is an open question.

CHAPTER 4 : THE REACTIVE PARADIGM

• Brooks criticized Hierarchal paradigm for it has a horizontal decomposition. From ethological literature, she suggested vertical decomposition. For primitive initial design, evolution of new layers, inhibition of older's, creation of parallel tracks.
• If anything happens to an advanced level, lower level operates.
• avoid collision - wander - explore - build maps
• When multiple behaviors are active
• Brooks use Subsumption architecture
• Arkin & Payton use potential fields.
• Most behavioral implementations include perceptual & motor schema, although not stated explicitly.
• PLAN is dropped, SENSE-ACT is tightly coupled to behaviors.
• S-A alone does not specify how behaviors are coordinated. This topic is what architectures deal with.
• S-A organization of behaviors:
1. Behaviors are independent
2. Each have distinct sense-act (perceptual-motor)
• In later implementations, instead of one sensor-one behavior, multiple-sensors for one behavior are adopted. But sensor fusion should be local to the behavior.
  
• Characteristics:
• Should execute rapidly.
• No memory, fixed action patterns are welcome. BUT should be based on innate releasing mechanisms.
• 5 characteristics of reactive paradigm:
1. Agent should be situated
2. Behaviors are basic blocks, overall behavior should be emergent.
3. Only local, behavior-specific sensing is permitted. This implies robot-centric coordinates.
4. Modular behaviors, decomposition of tasks should be designed in software engineering principles.
5. Motivation from animals welcome.
• Programming based on behaviors is elegant since:
• Behaviors are inherently modular, can be isolated tested easily
• Supports for incremental expansion
• Re-usability
• A reactive architecture should provide mechanisms of
• Triggering behavior
• Coordination of behaviors
• 3 architectures fir in Reactive Paradigm. Best formalized
1. Subsumption
2. Potential Fields
3. Rule Encoding

• Brook's architecture, popular because they are the first who walk, avoid, climb over obstacles without thinking, when compared with Shakey.
• Modules are Augmented Finite State Machines (AFSM) with registers, timers, and interruption mechanisms.
• Behaviors are released in stimulus-response way.
• 4 aspects:
1. Layers of competence (or intelligence): lower layers are for basic survival functions, where higher levels create goal-directed actions.
2. Behavior layers operate concurrent and independent, to handle conflicts modules in a higher layer may override/subsume output from lower layers, winner always higher layer.
3. Minimal usage of internal states
4. A task is accomplished by activating corresponding layer, but giving different tasks is not trivial.

• Level 0: Move forward, not collide
• From sonar: If any collision, stop, according to generated repulsive forces, turn and run away.
• Stopping prevents from side sweeping the obstacle while turning.
• Level 1: Wander
• A random heading from WANDER module, and sonar readings are inputs of AVOID module, it subsumes to output of RUNAWAY module in the Level 0 (inhibition mechanism).
• Level 2: Follow Corridor (in Brook's original, Explore)
• LOOK module identifies corridor, STAYINMIDDLE module finds direction and subsumes the output of WANDER. When LOOK module does not give direction (since computationally expensive), motion info from shaft encoders are employed by STAYINMIDDLE to compute new vector. If LOOK is broke down, robot moves according to only INTEGRATE, according to internal states, which is dangerous. A time constant, on supression and inhibition should be placed (for situations of no update for long time).
• Discussion:
• No need to change lower layers when adding new layers, a good software engineering .
• Robustness when higher levels are disabled somehow.
• Since no memory, robot may wander continuously in same region.

Potential Fields

• Vectors are always used to represent behaviors, vector summation is employed to combine behaviors.
• Motor action should be represented as potential field.
• Length of the arrow -> magnitude
• Array is viewed from bird's eye.
• Perceivable objects exert a force field on the surrounding space.
• Basic potential fields are 
• uniform
• perpendicular
• attraction
• repelsion
• tangential
• Attractive fields are useful to represent a taxis or tropism.
• Magnitude profiles (not constant magnitudes due to distance change, not binary) enable the robot not to behave jerky, represents reflexivity.
• Constant magnitude, linear drop-off and exponential drop-off.
• A primitive potential field is represented by a single function.
• Local minima is the most known problem, fields may sum to 0.
• For multiple sensors case, when there is a big obstacle, RUNAWAY behavior is applied for each sensor. This is where the direct coupling of sensing to action works. Data coupling is good since program will function good to accomodate more IR sensors. If vector sum is done internally, solution would be specific to robot (it is not elegant from the software engineering perspective).
• When an object in close vicinity of the robot encountered, since the output of a COLLIDE behavior could not stop robot, by only summing up of behavior vectors, not possible.. So panic and halt procedures applied in emergency.
• Selective attraction field, when docking to a station with specific position and orientation is required. Tangential field plus selective attraction enables robot, who approaches from behind. 3 perceptual schemas are required for docking:
1. relative direction of dock (when not visible)
2. recognize dock (for tangential field)
3. recognize front (for selective attention)
• Finite State Machines might be used to coordinate and sequence behaviors (3 perceptual - 3 motor schemas). Additionally, behaviors may be parametrized.
  
• Advantages of potential fields:
• Easier to visualize,
• Can be parametrized,
  
• 2D plane can b extended to 3D easily.
• NaTs (Navigation Templates) are employed by Mark Slack to overcome the local minima problem.
• Another solution to local minima problem is to insert a random vector field.

CHAPTER 5 : DESIGNING A REACTIVE IMPLEMENTATION

• Reactive robotics is frustrated from 2 important deficits:
• Design of behaviors is more an art than science.
• Integration of well-designed behaviors is a very challenging subject. Emergent behaviors etc.<>
• Many applications are best thought as a sequence of behaviors.  For example picking up and disposing a soda can.
• In this chapter, OOP is introduced based on the schema theory.
• Establishing the ecological niche of a reactive robot is important.
• 2 techniques for managing behavior sequence: ( Both are similar to IRM - Innate Releasing Mechanism )
• finite state automata
• scripts
• Program a small set of generic behaviors rather than designing a large set of specialized behaviors..
  

Behaviors as  Objects in OOP

• Schema theory is used as a bridge between concepts in biological intelligence and robotics.

• Schema is a class. Perceptual schema, motor schema and behavior are all derived from schema class.

• Schema class contains data structures and other schemas. Optionally, a coordinated control program exists, to control any methods of derived classes.

• Behaviors are composed of at least one perceptual schema, one motor schema. These classes act as methods for the Behavior class.

• Perceptual schema contains at least one method, which transforms sensor data into percept data structure.

• Motor schema contains at least one method which transforms percept data structure into a vector (action).

• Perceptual schema is linked to the sensors, while Motor schema is linked to actuators.

• Behavior may be composed of several Perceptual schemas, Motor schemas and even behaviors.

• Primitive Behavior is composed of only one PS and one MS.

• Abstract Behavior is assembled from other behaviors or have multiple PS and MS.

• Example : A primitive move-to-goal behavior

  • The objective is to place right colored trashes into right colored bins.

  • move-to-goal(goal-color) is the basic behavior. [Better than move-to-red and move-to-blue]
    

  • perceptual schema is extract-goal (goal-color). - A Method

  • motor schema is pfields.attraction(goal-angle, goal-strength). - A Method

  • percept data structure has goal-angle and goal-strength fields.

  • Affordance of color is used to extract where the goal is in the image.

  • The affordance of the Coke is the color while the information extracted from perception is the angle and size. This information (percept) is given to motor schema to move.

• Example : An abstract follow-corridor behavior : Different class diagrams and structures to create same behavior.

• Where do releasers go in OOP?

  • Perception serves two purposes: to release a behavior and to guide it.
    

  • The releaser itself may be a perceptual schema and behavior is instantiated when releaser condition is satisfied OR the behavior is always active and when releaser condition is not satisfied coordinated control program short-circuits processing.

  • The releaser must be designed to support correct sequence.
    

Steps in Designing a Reactive Behavior

1. Describe the task.

2. Describe the robot. The designer is usually handed some fixed constraints on the robot platform which will impact the design.

3. Describe the environment.

4. Describe how the robot should act in response to its environment (behavioral decomposition ecological niche).

• Behavior table is constructed which includes, releaser, Behavior name, Motor Schema, Percept, and Perceptual Schema
  

5. Implement & refine each behavior.

6. Test each behavior independently.

• It is important to remember that the simulators often only model the mechanics of the robot, not the perceptual abilities.
  

7. Test behaviors together.

In most simulation environments, motor schema is simulated correctly, however the simulation of perception and sensing mechanisms is not realistic.

Case Study : Unmanned Ground Robotics Competition

• In this section, the design steps of reactive behavior is realistically implemented.

Assemblages of Behaviors

• Concurrent run and in a sequence, how to formally define? The coordinated control program member of the abstract behavior expresses the releaser for the component behaviors.

• Most important point to remember is to make the world release, rather than rely on internal model.
  

• Finite State Automata :

  • states

  • inputs or alphabets : releasers

  • transition defines the new state due to current state and releaser

  • Obstacle avoidance should be always on and is implicit in FSA.

• A Pick Up the Trash FSA

  • Behavior table does not show the sequence and control of behaviors, so FSA can be used.

  • A common mistake is done when all releasers are not shown in FSA for the sake of previous state should have that releaser.

  • It is difficult to show the concurrency of behaviors using FSA.

• Implementation Examples

  • Schema-theoretic implementations use potential field methodologies and vector summation for effector's control.

  • Advantage of FSA is that they are abstract and can be implemented in a number of ways.

  • Inhibition and suppression cannot be well represented using FSA.

• Sometimes, robots fail in their task, these events are called exceptions. Another behavior or other parameters can be substituted in such situations.

• Scripts : ????

CHAPTER 7 : THE HYBRID DELIBERATIVE/REACTIVE PARADIGM

• Reactive paradigm by the end of 1980: real-time environments fast, inexpensive processing. No planning or reasoning of global state of robot. No general task accomplished, no performance monitoring. No making map. Cognitively oriented functions.
• Emergence of behaviors.
• Problem of 90's: put deliberation (planning) back into the robots without disputing the success of reactive behavioral control.
• Behavioral control in low level.
• 5 different architectures in this chapter:

1. AuRA (Arkin's)
2. Sensor Fusion Effects (SFX)
3. 3T
4. Saphira (more top-down)
   
5. TCA (more top-down)
   
6. Rhino ? & Minerva will be discussed in later chapters.
   

• Reactive planning.
• At first : for unstructured environments: reactive for knowledge rich environments : hybrid (which can be easily modeled)
• Now : Hybrids are the best general purpose architectures and solutions
• The use of asynchronous processing techniques.
• Good software modularity allows subsytems to be mixed effectively.
  

7.2

• Plan then Sense-Act.
• PLAN: all deliberation and global world modeling.
• 2 assumptions:
• deliberation works with symbols. Does not make 'fine-grained' decisions.
• Since expensive, decoupled from standpoint of practicality.
• Sensing in hybrid is complex. For behaviors, sensing is local and behavior specific. However, model making is global, so not behavior specific sensing.
• Percepts that model making processes use, might be created by perceptual schemas of behaviors or might be created by standalone observations.
• In hybrids:
• skill term is generally used instead of behavior since not only reflexive.
• instead of basic behaviors, use assemblages of behavior sequences over time
• more diversity in methods for combining outputs of concurrent behaviors.
• In hybrid, the term global not only used for modeling, but behavioral management, performance monitoring etc.
  
• Behaviors would execute until the plan was completed then planner generate new set of behaviors.

• Distinction based on three questions:
• How architecture distinguish between reaction & deliberation?
• Module functionalities
• Module access to global
  
• What is global knowledge to be incorporated
• How organize responsibilities in the deliberative portion?
• decomposition..
• How overall behavior emerges?
• How to combine reactive behaviors?

1. subsumption
2. potential fields
3. voting (DAMN)
4. Fuzzy logic (Saphira)
5. Filtering (SFX)

• Common components of hybrid architectures:
  • Sequencer: generates a set of behaviors
  • represented as a dependency network or FSM (dynamically generated)
    
  • Resource Manager: Allocates resources to behaviors, selects from libraries of schemas (ie. select appropriate among vision, sonar, IR)
  • Cartographer: May imply global world model, knowledge representations or map
  • Mission Planner: Human interaction and Make plans
    
  • Performance Monitoring
• Styles of hybrid architectures ( a loose division) :
  • Managerial styles : subdividing the deliberative portion into layers based on the scope of control or manageral responsibility.
  • State hierarchies : use knowledge of robot state to distinguish between reactive and deliberative activities.
  • Model-oriented styles : Characterized by behaviors that have access ro portions of the world model. turn out to appear to Hierarchical .....

• Concentrated on decomposition of responsibilities.
  
• From top to bottom:
• High level planning
• Plan is taken, resource is gathered
• Reactive behaviors
• Higher layers eavesdrop lower, and give directives
• Each layer carry out its directive locally. If not successful, ask help from superior, called fail upwards.

• oldest, by Arkin, based on Schema theory, 5 subsystems:
• Deliberative portion: Planner and Cartographer
• Reactive portion: Sensor and Motor subsystem
• between portions: Homeostatic Control
• Planner components interact with Cartographer (which is responsible from map making and reading):
• Mission Planner: Interface with human
• Navigator: Work with Cartographer, compute and segment paths
• Pilot: Take subtasks, gets relevant information to generate behaviors. Interacts with Motor Schema Manager. give list of behaviors to accomplish current subsegment. A number of motor schemas which coupled with perceptual schemas.
• Motor schemas are limited to potential fields.
  
• Schemas can consist of assemblage of schemas, and can contain specific knowledge and memory.
• Homeostatic Control  locates between Deliberative and Reactive layers. It changes gain due to the health of robot or other constraints. Similar to increasing hunger of animals.
• There is a link between Cartographer and Sensing subsystem.
  

• Started out as an extension to AurA by Murphy, to incorporate sensor fusion and handle sensor failure.
  
• It is an example how robustness can be built into an architecture.
• Motivated from sensing of cats:  each receptive field has a unique sense, then senses are streamed to both Cerebral Cortex (more cognitive) and Superior Colliculus (motor behaviors)  from receptive fields.
  
• Deliberative component divided into self-contained modules, which are specific to some area of competence.
• Dominant agent, Mission Planner, is responsible from human interaction, and specification of mission constraints.
• The agents in lower deliberative layer get direction from Mission planner, and give directions to workers (behaviors) in Reactive layer.
  
• Resource Managers are: Task Manager, Sensing Manager, and Effector Manager use AI planing, scheduling, and problem solving for best allocation of effector and sensing resources.
• Sensing Manager deals with failures in behaviors (ie. perceptual schema). There is an example of shaking its camera, when camera is covered with t-shirt.
• Cartographer agent & performance monitoring agents are also in lower deliberative layers for map construction and observation of progress towards goal, respectively.
• Reactive component:
• SFX use filtering method for combining outputs. Similar to subsumption, but reverse, lower layers subsume higher layers. Lower corresponds to Tactical, and higher corresponds to Strategical.
• Why tactical-strategical? Tactical behaviors serve as filters on strategic commands to ensure the robot acts in safe manner, as close accordance with the strategic intent as possible

• predominantly used by NASA, contains aspects of Slack's NAT system, Gat's subsumption-style ATLANTIS, and Firby's RAPs systems.
• One deliberative, one reactive and one interface between two.
• Primary usage on planetary rovers, underwater vehicles, and robot assistants to astronauts.
• Top layer: Planner fulfills duties of mission planner and cartographer.
• Sequencer located at middle layer, goals are passed to it. Use reactive planning technique RAPS, to develop a task network.
• Controller or Skill Manager at lowest layer contains skills which are assemblage of primitive skills.
• Skills are associated with events. Events serve as checkpoints on success of actions.
• Algorithms that have low update rates are put higher layers (ie. Planner).
  

• A more top-down, not behavior or skill based, have symbolic flavor. Symbolic manipulation around global world model.
• Unlike others (managerial or state-hierarchies), global world model supplies perception to behaviors (as virtual sensors).
• Four conceptual differences:
1. global world model less ambiguous
2. distributed perceptual processing ? why different ?
   
3. sensor fusion over time ? why different?
4. increase in processor speed

• Used on direct descendants of Shakey: Flakey and Erratic by Kurt Kanolige at SRI
• 3 key elements in mobile robotics:
1. coordination (of actuators, sensors and goals)
2. coherence (maintain global world)
3. communication (human)
• Deliberative layer:
• a type of reactive planner, PRS-lite, take voice commands, and produce navigation tasks.
• Local Perception Space (LPS) serves as central world model, is basis for planning.
• Independent software agents perceive world.
• Main emphasis is on construction and maintenance of global world model
• Behaviors get commands from Topological Planner and Navigation Tasks on Deliberative level which depend on LPS.
• Reactive Layer:
• Input to reactive behaviors are from virtual world model (LPS)
• Fuzzy logic is employed in coordination of behavioral outputs.

• used in NASA, designed by Reid Simons, applied on Xavier
• Sensors are fed to Global World Model.
• Task Scheduling interacts with people, prioritize tasks
• Path Planning and Partially Observable Markov Decision Process (POMDP) use world model.
• Avoidance layer use evidence grids.

• Highly modular
• high degree of niche targetability
• emphasis on robustness
• Differences from Hierarchical:
• employment of software engineering principles
• no frame problem (in terms of execution)
• not plan every move in lowest level of granularity
• roots in ethology
  

• Managerial and State-Hierarchy Architectures are evolved from Reactive Paradigm. However, Model-Oriented paradigm is a more top-down approach. It contains more symbolic flavor than others. Supply virtual sensors from world models, into behaviors.
• A specific Model-Oriented Architecture : Task Control Architecture (TCA). Extensively used in robots designed by NASA, especially for service robotics. Figure, 7.10.
  • Task-Scheduling Layer: interact with user, prioritize task and optimize schedule.
  • Path Planner : Plan the route that robot should follow.
  • Navigation : The next stop, or node which should be gone.
  • Obstacle Avoidance : reflexive movements.