AI researchers vs control engineers
- AI res: solve problems by increasing the autonomy (e.g. do the the right thing in an open world approach)
- Control eng: solve problems by extending the automation (e.g. modelling the world in order to create a closed world and guarantee “effect of actions”)
They are not necesseraly mutually exlusive.
The 3 types of software architectures
- Operational Architecture: description of what the systems does (high level)
- Systems Architecture: describes the major blocks (or subsystems) that make the systems works (mid level)
- Technical Architecture: describes the details of the implementation (low level)
Why the 3 architectures are useful for programming a robot?
Encoraunges the creation of designs that meet the four general principles of software engineering:
* abstraction: no details
* modularity: object-oriented programming
* possibility of substituting old modules
* generality/portability: allowing the code to be used for other applications
What are the 4 primitives used in robotics?
- SENSE: sensor data -> sensed information
- ACT: sensed information or directives -> actuators commands
- PLAN: informations -> directives
- LEARN: mechanism for maximizing changes of success (it trascend the other primitives)
Canonical AI robot operational architecture (layers and attributes)
AI robotics has converged on the hybrid deliberative/reactive operational architecture.
Organized in three layers (inspired by the biological intelligence):
- reactive
- deliberative
- interactive
Attributes of the layers:
- primitives: sense, plan, act, and learn
- perceptual ability: direct or recognition
- planning horizon: present or past - present or past - present - future
- time scale: fast response or slow (planning)
- use of models
Robot’s behavior and types
Refers to a specific task or action that a robot can perform in response to a specific set of input or conditions.
- reflexive behavior: hard-wired for fast response (kneed-jerk reaction)
- reactive behavior: learned and compiled down (playing piano)
- conscious behavior: require deliberative though (writing code)
Reactive layer of the canonical operational architecture
- SENSE + ACT
- Corresponds to the reactive loop in the central nervous system. This loop is associated with functionality that occurs in the spinal cord and “lower brain”, especially responses and skills based on motor memory.
- No awareness of the larger world.
Deliberative layer of the canonical operational architecture
PLAN (using a world model) -> SENSE-ACT
Corresponds to the cognitive loop associated with the cortex in the brain. The brain independently takes as inputs the same signals used by the “lower brain” and adds additional sensor processing to make conscious decisions.
Cognitive loops can modify reactive loops.
Speed and content are different: this can explain why the two loops are asynchronous (e.g. of the office worker that being the falling cup of coffee).
The 4 deliberative functions
- generating: planning, reasoning and problem solving in AI
- monitoring: anticipate potential failures
- selecting: resources allocation
- implementing: execution of the plan
Interactive layer of the canonical operational architecture
Need for intercation with other agents, either people, other robots, or software agents.
What are the 3 types of sensing routes
- Global sensing: many-to-one relationship because we have a function that merges the sensed data into a World Model
- Local sensing: one-to-one or one-to-many because we have the sensed data that goes into indipendent functions
- Hybrid sensing: combination between the two above ones
The 3 Systems Architecture paradigms
- Hierarchical
- Reactive
- Hybrid deliberative/reactive
They arrange the flow of data and control within the robot.
Each paradigm is uniquely described using two traits: the pattern of interaction between the 3 robot primitives and the route of sensing
What are the 5 subsystems in Systems Architectures?
- Planning: generating and monitoring overall mission objectives
- Navigation: generating and monitoring paths of movements and selecting the resources to accomplish the movements
- Carpographer: construction and maintenance of knowledge representations about the world (world model)
- Motor Schema: selecting the best motor routines and implementing actions
- Perceptual Schema: selecting the best sensors for the motos actions and implementing sensor processing algorithms
draw the schema 4.6
2 important aspects of the 5 subsystems
- they are not independent and do not represent a sequence of programming actions. The “main” program employs subsets of their functions and data structures to produce capabilities
- the list is not necessarily complete
Hierarchical systems architecture paradigm
- relies of global sensing routing
- used where the mission or application is well-understood
SENSE-PLAN-ACT (cycle)
- SENSE: Perception subsystem and Cartographer subsystem
- PLAN: Planner subsystem and Navigation subsystem
- ACT: Motor Schema subsystem
draw the schema 4.9
Advantages and disadvantages of the hierarchical systems architecture paradigm
ADVANTAGES:
- simple ordering between the primitives (intuitively appealing and consistent).
- optimizes programming
DISADVANTAGES:
- planning is very slow
- closed world assumption (frame problem)
Reactive systems architecture paradigm
SENSE-ACT: only 2 subsystems, Perception and Motor Schema
- it uses in parallel multiple instances of the Perception and Motor Schema subsystems
- born because of the dissatisfaction with the results of the hierarchical paradigm
- users loved it because it worked but AI people wanted to put PLAN back in
- control people hated it because couldn’t rigorously prove it worked
Examples of behaviors for the reactive systems architecture paradigm
- exploration: general direction
- goal-oriented: towards an attractor
- avoiding: prevent collions
- path following: move on a designated path
- …
First reactive architecture
Subsumption architecture.
It consists of a hierarchy of layers, each containing reactive behaviors. The lowest layer contains the most basic behaviors, such as avoiding obstacles or following a path. These behaviors are executed continuously and have priority over the behaviors in the higher layers.
Other technniques can be applied…
Who is the arbiter and what is the arbitration?
- arbiter or orchestrator: A very specialized behavior module monitoring robot operation which selects the most suitable behaviors to be activated depending on new acquired sensor data.
- arbitration: the implemented activity of the arbiter which results in either spatial or temporal ordering of robot behaviors
Foraging
- a robot moves from its base looking for objects that attract it (wandering and avoiding obstacles)
- when it finds an object, it moves toward it, pick it up (picking), and returns to the base (homing)
- repeats the sequence until all objects have been brought to the base
draw the subsumption hierarchy
Other techniques for selecting a behavior in the subsumption architecture
- competitive: only the output of one of the behaviors is selected
- cooperative: it blends the outputs of different behaviors in some way
- mixed
Hybrid deliberative/reactive systems architecture paradigm
- relies on both global and local sensing routing
PLAN while SENSE -> ACT
Similar to the Hierarchical but the sensored data is splitted and given to the Cartographer and directly to the Motor Schema.
Can intelligence be added in layers?
Yes: a robot may not need all the layers for a particular task.
No: e.g. a successful auotonomous capability such us Navigation needs all the layers.
