9 - Control Architectures Flashcards
Name 5 problems you may encounter with robotic behaviours?
Complex environment
Noise/unpredictable environment
Actions don’t always lead to intended consequences
Misleading sensors
Responses are too sloe
Why do you need control architectures?
How to bring a range of competencies together into a single system that can operate autonomously (and reliably, in a complex world)
What is the definition of the robot control paradigm?
A philosophy or set of assumptions and/or techniques which characterise an approach to a class of problems
What are the 3 main control architecture paradigms?
Deliberative, reactive, hybrid
How do you typically characterise the 3 control architecture paradigms?
Through the fundamental primitives, sense, plan, act
What is sense?
Sensing the environment, takes raw sensor data and returns the sensed information
What is plan?
Deciding what to do using some model of the world, takes sensory information and returns directives
What is act?
Acting on the environment, takes in sensory information of directives and returns actuator commands
What is the extra control architecture?
Cognitive
What is a deliberative architecture?
Planning what action to take, assuming you have a world model then doing it.
What is the basic process of a deliberative architecture?
Gather currently available information and integrate it into the world model, plan what to do, then execute the plan and return the step 1
What does the deliberative architecture emphasise?
A top down (hierarchical) planning process, partly inspired by human introspection
What is STRIPS?
Stanford Research Institute Problem Solver
Symbolic representation of all information
What are the 4 parts to STRIPS?
The world model (everything about the state of the environment)
The capabilities/properties of the robot itself (operators)
Initial and goal states
Difference evaluator (how close to the goal state am I?)
What is PDDL?
Planning Domain Definition Language.
STRIPS plus extensions, common assumptions, benefits, shortfalls
What is the difference between STRIPS and PDDL?
STRIPS is a specific planner/language, PDDL is a more recent standardised planner
Name 5 limitations of deliberative architectures?
Closed World problem (all information is present, nothing unexpected, no unanticipated consequences)
Frame problem (what is and is not relevant, should enumerate all states even if unchanged)
Brittleness problem (can’t handle change not effected by the agent)
Uncertainty problem (how should this be handled in a symbolic planner that assumes crisp knowledge and true/false conditionals)
Computational load (high load leads to slow reactivity)
What is a reactive architecture?
Direct reaction against deliberative models
Emphasis on fast reaction to low-level sensory information, without involved processing and planning
What is the basic process of reactive architectures?
Sensory input acquired, multiple parallel behaviours result in overt agent actions
What are 3 benefits to reactive architectures?
No internal world model is needed
Real-time behavioural control
Can have emergence of complex behaviour with little design effort
What is behaviour based robotics?
Typically reactive, tightly coupled to sensor information (no planning or internal state)
In the design of the robots, strong focus on robot’s body
Interaction with environment
What are ego-centric representations?
Deals with the limitation of having no world model.
Provides some structure and helps with choosing what to do
Name 3 roles for a coordinator in multiple behaviours?
Competitive (winner takes all)
Cooperative (blend outputs through addition)
Hybrid (activation/inhibition dynamics)
What is a subsumption architecture?
A single example case of behaviour-based control architecture
Gets around the coordinator problem by having higher level behaviours subsume lower level behaviours (i.e some behaviours can over-ride others)
What are the limitations to reactive architectures?
Oriented to specific task (lack of generalisation)
Based on, and constrained by particular robot embodiment
Sensitivity to sensor noise
Lack of planning
Stimulus response alone insufficient to account for intelligence
Emergence of complex behaviour is a design problem
What are hybrid architectures?
Trying to get the best of both deliberative and reactive paradigms
Multiple levels of control, each focused on a difference aspect
Some planning where appropriate but maintaining ability to respond quickly to the environment
How can you get the best of both in hybrid architectures?
Use symbolic processing and world models/maps for planning and use reactive behaviours for fast responsive action
What is temporal decomposition?
Deliberative uses horizontal decomposition of the task and reactive uses vertical decomposition of tasks, temporal resolves this by using time-appropriate processes where different layers have different speeds of processing
Discuss the speed of horizontal decomposition and vertical decomposition
Horizontal decomposition of the task is used in deliberative architectures and can be slow, whereas vertical decomposition of the tasks are used in reactive architectures and can be too quick
What are the 3 tiers in the Three Tier (3T) architecture?
Controller, sequencer/executive/deliberator
Discuss the controller in 3T
Library of behaviours, must be fast, must be able to detect failure (allows sequencer/executive to call another behaviour or call for replanning)
Discuss the sequencer/executive in 3T
Drives the control of the system, selects which behaviour will be active, initiates behaviour and planning
Discuss the deliberator in 3T
Operates on its internal state (world model, no sensing), time consuming and computationally intensive tasks, no commitment to the means of processing, and operation is directed by the sequencer/executive
What are the advantage of 3T and hybrid?
Specification of overall structure, and principle of operation rather than precise mechanisms
Maintains flexibility
If one algorithm is not appropriate then swap it out for another
What are the issues with hybrid architectures?
Central role of the sequencer/executive
Symbol grounding (relevance of the representations to the instantiation/actions of the robot system it is planning for)
What is a cognitive architecture?
Explicitly takes into account the way that humans may process information and act
Overlaps with the previous paradigms, particularly hybrid
What is the definition of a cognitive architecture?
The overall, essential structure and process of a domain-generic computational cognitive model, used for a broad, multiple-level, multi-domain analysis of cognition and behaviour
Why take inspiration from humans for cognitive architectures?
Humans demonstrate the best example of highly complex intelligence, and so could form useful design guides
If robots are to interact with humans in human environments then having them endowed with some human-like cognitive features could be useful
What are 2 main approaches to cognitive architectures?
Derive a set of mechanisms to use from human behavioural or other data
Try to model fundamental principles of organisation of cognition and implement these
What is ACT-R?
An established cognitive architecture that has been applied to numerous models of human behaviour
Based on evidence from pscyology and pysiology
Hybrid symbolic/sub symbolic processing
What is HAMMER?
Hierarchical, Attentive, Multiple Models for Execution and Recognition
Fundamentally based on forward/inverse model couplings, and applied to imitation, learning, assistance etc
What are 3 issues with cognitive architectures?
How related to biology should it be? (inspiration or constraints)
Suitable level of abstraction? (behaviour or mechanism)
What is the purpose of using a cognitive architecture? (functional or explanatory)
What are multi-agent systems?
Coordinating multiple robots, emergent behaviour from swarms of robots
