Architectures for Autonomy PDF

4a Architectures for Autonomy Given that autonomy has a different programming style, what is it? How Much Artificial...

4a Architectures for Autonomy Given that autonomy has a different programming style, what is it? How Much Artificial Intelligen ce Does a Need? Robot Can intelligence be added in layers? Like upgrading to “pro version” or downloading “apps” as needed? © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) http://www.gameartisans.org 1 4a Specific Learning Objectives Apply Levis' definition of architectures to organizing software in an intelligent robot and be familiar with how they can guide the design process Objectives Review Architecture Name and describe the layers (behavioral, deliberative, -types Operational - Biological interface) within a canonical operational architecture of an -Process intelligent robot in terms of the five attributes -Functions -Ramifications Layered Int. How much? Name the four primitives of robotic intelligence (sense, plan, Summary act, learn) List vulnerabilities (e.g., where poor choices or implicit assumptions can lead to failures) in the canonical operational architecture © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 2 4a Objectives GIVEN THAT AUTONOMY HAS A DIFFERENT PROGRAMMING Review Architecture -types Operational STYLE, WHAT IS IT? - Biological -Process -Functions -Ramifications Layered Int. How much? Summary Via… Architectures, specifically operational architecture Biological operational architecture Recast as a Process Recast as a Function Ramifications © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 3 4a Organizing Software Overall style of design or organization is called an Objectives Review “architecture” Architecture -types Operational - Biological – provides a principled way of organizing a control system. -Process -Functions -Ramifications However, in addition to providing structure, it imposes Layered Int. How much? constraints on the way the control problem can be solved Summary [Mataric] – describes a set of architectural components and how they interact [Dean & Wellman] © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 4 4a Types of architectures [Levis, George Mason University] Objectives operational architecture: describes what the systems does at a high level, not how it does it Review Architecture -types Operational - Biological -Process -Functions -Ramifications systems architecture: describes how a system works in terms of major subsystems Layered Int. How much? Summary technical architecture: describes how a system works in terms of implementation details, language © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 5 4a Types of architectures [Levis, George Mason University] Objectives operational architecture: describes what the systems does at a high level, not how itBuyer does itpicks out Review Architecture -types Operational - Biological -Process plans for a -Functions -Ramifications systems architecture: describes how a system house works in terms of major subsystems Layered Int. How much? Summary Contractor lines up plumbers, technical architecture: describes how a system roofers… works in terms of implementation details, language Studs go here, pipes here… © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 6 4a Focus of this Course Objectives Review Architecture operational architecture: describes what the systems does, not how it does it -types Operational - Biological -Process -Functions -Ramifications Layered Int. How much? systems architecture: describes how a system works in terms of major subsystems Summary technical architecture: implementation details, language © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 7 4a Often Changes with New Languages, Applications Objectives Review Architecture operational architecture: describes what the systems does, not how it does it -types Operational - Biological -Process -Functions -Ramifications Layered Int. How much? systems architecture: describes how a system works in terms of major subsystems Summary technical architecture: implementation details, language Subject of Standards Wars © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 8 4a Why Worry About Software Organization? Objectives Recall at least 7 distinct areas of Artificial Review Architecture -types Intelligence, each with own algorithms and data Operational - Biological structures; these have to be “knitted” together -Process -Functions -Ramifications somehow Layered Int. How much? Summary Software engineering is necessary for a successful software enterprise © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 9 4a Thinking About Architectures is Good Software Engineering Abstraction – Ignores details to permit focus for thinking about general organization of Objectives intelligence Review Architecture – Semiformal description -types Operational - Biological -Process Modularity -Functions – high cohesion (do one thing well, may be able to substitute different -Ramifications Layered Int. algorithms which produce same results “logical sensors”, “logical How much? Summary behaviors”, etc) – low coupling (which means may be able to add “apps”) – Supporting unit testing and debugging Anticipation of change, Incrementality – How to adapt, support evolution Generality – Not re-invent the wheel each time © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 10 4a Canonical Architecture Objectives Review “Upper brain” or cortex Architecture Reasoning over symbols Interaction -types Operational (information) about goals Layer - Biological -Process -Functions Deliberative -Ramifications “Middle brain” Layer Layered Int. How much? Converting sensor data into Summary symbols (information) Reactive (or Behavioral) Spinal Cord and “lower brain” Layer Skills and responses © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 11 4a Canonical AI Robotics Operational Architecture – Attributes for Describing Layers – The description and justification for what software functions go into each layer are based on five attributes: – 1-primitives – 2- perceptual ability – 3- planning horizon – 4- Time scale – 5- Use of models. © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 12 4a Primitives for Robot Intelligence Objectives Review Architecture -types Operational - Biological SENSE PLAN ACT -Process -Functions -Ramifications Layered Int. How much? Summary LEARN © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 13 4a © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 14 4a Consider as a Process: OODA Objectives Review Architecture -types Operational - Biological -Process -Functions -Ramifications Layered Int. How much? Summary OODA loop, from military pilots – Observe, Orient, Decide, and Act – Problem: not everything intelligent requires “deciding” (ex., reflexes or muscle memory such as riding a bike or driving a car) You Tube video on John Boyd and OODA http://www.youtube.com/watch?v=ivTBv3wnp1Yt © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 15 4a Consider as a Process: AOP Objectives Review Architecture -types Can be Operational - Biological Null -Process -Functions -Ramifications Layered Int. How much? Summary Action-Perception Cycle from cognitive science, artificial intelligence – Action-oriented perception is similar to OODA stronger scientific basis and acceptance “short cuts” for reflexes, reactive behaviors – AI has cast these as 4 components or primitives: SENSE, PLAN, ACT, and LEARN – Problem: too high level to be of much value in software organization; for example different complexities of what to sense (ex. proximity to an obstacle vs. obstacle is unusual for this situation and indicates an ambush) © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 16 4a Canonical Architecture` Interaction: Persona and teaming Objectives Review Deliberative loop: Interaction Architecture Reasoning over symbols -types Layer Operational - Biological -Process Key function: -Functions Converting sensor Deliberative -Ramifications Layered Int. data into symbols Layer How much? Summary Reactive (Behavioral) Reactive loop: Skills and Layer responses © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 17 4a What Does This Mean for Programming? From Reactive to Deliberative Objectives Review – Two types of perception: DIRECT, RECOGNITION (symbols) Architecture -types Impacts computer vision Operational - Biological – Different time horizons -Process -Functions From Present to Present, Past, Future -Ramifications Layered Int. Impacts sensing, storage, as well as algorithms reasoning, projecting How much? Summary – Need a central structure (WORLD MODEL) to hold the symbols, history, knowledge but is tractable From Reactive/Deliberative to Interaction – Additional knowledge “theory of mind”– beliefs, desires, intentions (BDI) of the other agent, common ground © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 18 4a The Hard Parts From Reactive to Deliberative Objectives Review – Two types of perception: DIRECT, RECOGNITION (symbols) Architecture -types Impacts computer vision Operational - Biological – Different time horizons -Process -Functions From Present to Present, Past, Future -Ramifications Layered Int. Impacts sensing, storage, as well as algorithms reasoning, projecting How much? Summary – Need a central structure (WORLD MODEL) to hold the symbols, history, knowledge but is tractable From Reactive/Deliberative to Interaction – Additional knowledge “theory of mind”– beliefs, desires, intentions (BDI) of the other agent, common ground © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 19 4a Objectives Review Architecture -types Operational - Biological -Process -Functions -Ramifications Layered Int. How much? Can we make this more tangible? Summary © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 20 4a “Behavioral Robotics” Objectives Review Architecture -types Operational - Biological -Process -Functions -Ramifications Layered Int. sense act sense act How much? Reactive (or Behavioral) Summary Layer sense act “born” with sets of SENSE-ACT couplings called behaviors that get turned on/off based on stimulus No PLAN © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 21 4a Control Theory is “Lower Level” But Doesn’t Necessary Capture it All Objectives Reactive (fly by wire, inner loop Review Architecture control, behaviors): -types Operational - Biological Tightly coupled with sensing, so very -Process -Functions fast -Ramifications Layered Int. How much? Many concurrent stimulus-response Summary behaviors, strung together with simple sense act sense act scripting with FSA sense act Action is generated by sensed or internal stimulus No awareness, no mission monitoring Models are of the vehicle, not the “larger” world © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 22 4a Behavioral Robotics Focus in AI from 1986-1996 Objectives Things you can do with behavioral robotics Review Architecture – Roomba, Aibo -types Operational – Guarded motions (don’t hit anything), panic behaviors (stop!), - Biological -Process “Macros” such as self-righting in teleoperation -Functions -Ramifications Layered Int. How much? Advantages Summary – Direction perception is usually simple – High modularity, add new behaviors without reprogramming old working behaviors, degrades gracefully (depending on the technical architecture) © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 23 4a Behavioral Robotics Issues Hidden Costs – Behaviors plus some sort of coordination function to Objectives Review fuse outputs to effectors Architecture -types Operational - Biological Where it breaks -Process -Functions – “Fly at window” effects due to local scope (but lots of -Ramifications Layered Int. ways around this) How much? Summary – Poor choice of coordination functions (go left, go right = go in middle and hit obstacle) Where it doesn’t break but scares/annoys people – Since sensor/situated based often can’t predict whether will go left of right to avoid (but only if doesn’t matter, see NaTs) – Not optimal because sensor based (reactive) versus symbol/map based © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 24 4a “Hybrids” plan Objectives Deliberative World Review Layer plan Architecture -types model Operational - Biological -Process -Functions -Ramifications Layered Int. sense act sense act How much? Reactive (or Behavioral) Summary Layer sense act PLAN, then instantiate appropriate SENSE-ACT behaviors, until next step in plan, … PLAN requires a World Model (though it is bounded) plus the actual planning algorithms © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 25 4a 3 Tier Variant plan Planner Objectives World Review plan Architecture -types model Operational Sequencer - Biological -Process -Functions -Ramifications Layered Int. sense act sense act How much? Reactive (or Behavioral) Summary Layer sense act PLAN, then instantiate appropriate SENSE-ACT behaviors, until next step in plan, … Technical PLAN execution but a significant programming aspect © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 26 4a LEARN Different types of learning, different things to learn– is very diffuse so more permeates the architecture Objectives Review Architecture -types Operational - Biological -Process -Functions -Ramifications Layered Int. How much? Summary © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 27 4a So Far… Biological organization suggests three layers of intelligence with distinctly different perception, Objectives Review knowledge, planning horizons, and time scales Architecture -types Operational - Biological -Process The AI Robotics field has converged on PLAN, then -Functions -Ramifications SENSE-ACT with LEARN as needed at different Layered Int. How much? Summary points – Technically this is SENSE-PLAN, SENSE-ACT but historically the sensing for planning just like the execution monitoring is lumped in “PLAN” But still not as tangible as desired… © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 28 4a Consider as Functions Large scale factory automation (not factory robots) has 4 functions – generating – selecting Objectives Review – implementing Architecture -types – Monitoring Operational - Biological -Process May involve different contributions/combinations of primitives -Functions -Ramifications – Ex. monitoring requires sensing, understanding what was planned, and Layered Int. How much? perhaps learning of what is normal and not normal Summary Factory automation functions don’t capture implementation but does seem to capture deliberation – what was implemented was usually a single, monolithic routine or control loop; natural intelligence usually have many low-level “routines” running concurrently Problem: still missing data/knowledge structures; not prescriptive: if I want to do X, then pick these components or have this set of layers… © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 29 4a PLAN got Bigger and Matches 3T Upper level (PLANNER) is mission generation & Objectives monitoring generating monitoring Review Architecture Need Past, Present, -types World Future Operational plan - Biological -Process model Lower level -Functions selecting implementing -Ramifications (SEQUENCING) Layered Int. How much? is selection of behaviors Summary to accomplish task sense act sense act (instantiation) & local sense act monitoring Need Past, Present © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 30 4a Other Ramifications: Planning Horizon, Time Scales Objectives Review Architecture monitoring generating PRESENT+PAST+FUTURE, SLOW -types World Operational plan model - Biological -Process -Functions selecting implementing PRESENT+PAST, FAST -Ramifications Layered Int. How much? Summary PRESENT, VERY FAST sense act sense act sense act © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 31 4a Canonical Operational Architecture Objectives Review Architecture Interaction Each layer has a -types Layer Operational - Biological different style of -Process -Functions program organization -Ramifications Layered Int. How much? Deliberative Summary Layer Not concerned with Reactive Layer details of interaction layer Behavioral Layer © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 32 4a Programming Languages Objectives Review Architecture Interaction Procedural, Functional, -types Operational Layer Ontological languages such - Biological as OWL -Process -Functions -Ramifications Layered Int. How much? Deliberative Summary Layer Functional languages such Reactive Layer as Lisp Behavioral Procedural languages such Layer as C, C++, Java © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 33 4a Advantages of programming in layers with different styles Decomposition of a complex system Objectives Review Architecture – Can use a separate processor(s) for each layer or -types Operational behaviors - Biological -Process – Can split between on-board, off-board -Functions -Ramifications Layered Int. How much? Summary Matching right tools and mindset for the task – Ex. C++ for behaviors, Lisp for planning Add to working, verified code © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 34 4a Given that autonomy has a different programming style, what is it? Objectives Review Architecture -types Operational - Biological There are three different styles as formalized by the 3T hybrid operational architecture -Process -Functions -Ramifications Layered Int. How much? The styles reflect 5 attributes: Summary – Primitives involved (SENSE, PLAN, ACT) – Perception (direct or recognition) – Planning horizon (present, present+past, present+past+future) – Time scale (very fast, fast, slow) – Models (local, global) © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 35 4a Objectives Can intelligence be added in layers? Like upgrading to “pro Review Architecture -types Operational version” or downloading - Biological -Process -Functions -Ramifications “apps” as needed? Layered Int. How much? Summary © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 36 4a Can intelligence be added in layers? Like upgrading to “pro version” or downloading “apps” as needed? Objectives Review Architecture Yes, intelligence is organized in layers -types Operational - Biological Yes, Can create libraries of equivalent (“logical”) -Process -Functions algorithms -Ramifications Layered Int. How much? But as will be seen in later lectures… Summary – Adding new behaviors or algorithms is non-trivial Coordination functions at reactive layer impose certain assumptions, restrictions, and side effects – Adding another layer is non-trivial Different attributes such as perception, models require significant design investment “Hidden” coordination between layers © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 37 4a Objectives HOW MUCH ARTIFICIAL INTELLIGENCE DOES A Review Architecture -types Operational ROBOT NEED? - Biological -Process -Functions -Ramifications Layered Int. How much? Summary © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 38 4a How Much Artificial Intelligence Does a Robot Need? It depends… Objectives Review What functions does the robot need to do? Architecture -types – Generate? Monitor? Select? Implement? Execute behaviors? Operational - Biological Learn? -Process -Functions -Ramifications What planning horizon does the functions require? Layered Int. How much? Summary – Present, Present+Past, Present+Past+Future How fast do the algorithms have to update? – May have to use a closed world and guaranteed execution rates (control theory) What type of model does the robot need? Local? Global? Both? – Note: go with the minimum © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 39 4a Note that “How Much” Rating Systems Aren’t Tied to Programming Objectives Review Architecture -types Operational - Biological -Process -Functions -Ramifications Layered Int. How much? Summary Am not picking on ALFUS, it’s just the most recent attempt © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 40 4a Objectives Review Architecture -types Operational - Biological -Process SUMMARY AND ADDITIONAL -Functions -Ramifications Layered Int. How much? THOUGHTS Summary © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 41 4a Summary: Architectures An architecture is the Big Picture of how to program an intelligent robot Objectives Review “Architecture” refers to either Architecture – The operational architecture which described what the system does on an -types Operational abstract level - Biological -Process – The system architecture developed by a manufacturer (or research group) -Functions and may look like a data flow diagram -Ramifications Layered Int. – The technical architecture, specifies the actual techniques and code How much? organization Summary The canonical operational architecture for an AI robot consists of three layers which generally represent different programming styles or even implementation languages: – Behavioral – Deliberative – Interaction © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 42 4a Other Observations Current practice is good with deliberative functions operating on symbols, good with behaviors using Objectives Review direct perception Architecture -types Operational - Biological Major barrier is going from sensory data to symbols: -Process -Functions -Ramifications recognition and labeling as unique instances Layered Int. How much? – May be able to see a coffee cup but not that it is my coffee cup Summary and yours from the same dish collection is over there Major barrier in understanding human intention (which is often implied but never spoken); AI robots currently require explicit directions © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 43 4a Future Questions and Lectures Since some aspects seem beyond current Objectives Review programming capabilities, what are levels of Architecture -types Operational autonomy in terms of how much the robot can be - Biological -Process trusted to do and how much the human has to do? -Functions -Ramifications Layered Int. How much? Summary What is the systems architecture for the hybrid architecture? Where do the 7 areas of AI fit in with the operational architecture? © 2019 Robin Murphy Introduction to AI Robotics 2nd Edition (MIT Press 2019) 44