Last update: February 24, 2020 Chapter 1 Automated Planning Introduction and Acting Malik Ghallab, Dana Nau and Paolo Traverso Dana S. Nau http://www.laas.fr/planning University of Maryland Nau – Lecture slides for Automated Planning and Acting Licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License 1
Motivation Actor ● Actor : agent that performs actions Deliberation components Objectives Deliberation Other Planning Queries ● Deliberation functions components actors Plans Messages Acting ▸ Planning What actions to perform Commands Percepts ▸ Acting Execution platform Execution platform How to perform them Actuations Signals External World Nau – Lecture slides for Automated Planning and Acting 2
Planning ● Relies on prediction + search ● Different types of actions ⇒ ● Uses descriptive models of the actions ▸ Different predictive models ▸ Predict what the actions will do ▸ Different planning problems and techniques ▸ Don’t tell how to do them ● Search over predicted states and possible ▸ Motion and manipulation planning organizations of feasible actions ▸ Perception planning ▸ Navigation planning ▸ Communication planning ▸ Task planning a s ′ = γ( s,a ) s Most AI planning • • • Nau – Lecture slides for Automated Planning and Acting 3
Acting ● Traditional “AI planning” view: Actor ▸ Carrying out an action is just execution Deliberation components ▸ Can ignore how it’s done Objectives Deliberation Other Planning Queries components actors Plans ● Sometimes that’s OK Messages Acting ▸ If the environment has been engineered to make actions predictable Commands Percepts ▸ Example on next slide Execution platform Execution platform ● Usually acting is more complicated Actuations Signals ▸ Example later External World Nau – Lecture slides for Automated Planning and Acting 4
Acting as Execution Video: https://www.cs.umd.edu/~nau/apa/kiva.mp4 Nau – Lecture slides for Automated Planning and Acting 5
Deliberative Acting Video: https://www.cs.umd.edu/~nau/apa/crow.mov Nau – Lecture slides for Automated Planning and Acting 6
Deliberative Acting ● Actor is in a dynamic unpredictable environment ▸ Adapt actions to current context ▸ React to events ● Relies on ▸ Operational models telling how to perform the actions ▸ Observations of current state Planning stage Acting stage Nau – Lecture slides for Automated Planning and Acting 7
Planning and Acting ● Multiple levels of abstraction Actor ▸ Actors are organized into physical subsystems Deliberation components ▸ Deliberation reflects this Objectives Deliberation Other Planning Queries components actors Plans ● Heterogeneous reasoning Messages Acting ▸ Different techniques Commands Percepts • at different levels • different subsystems at same level Execution platform Execution platform ● Continual online planning Actuations Signals ▸ Can’t plan everything in advance ▸ Plans are abstract and partial External World until more detail is needed Nau – Lecture slides for Automated Planning and Acting 8
Bremen Harbor Nau – Lecture slides for Automated Planning and Acting 9
Example: Harbor Management manage incoming shipment ● Importing/exporting cars Planning ▸ Based on Bremen Harbor unload unpack store await order prepare deliver ● Multiple levels of abstraction … … … … … ▸ Reflect physical organization of harbor storage area A manager ● Continual online planning registration storage ▸ Top level can be planned offline storage area B … manager assignment ▸ The rest is online, based on current manager manager conditions storage area … C manager ● Heterogeneous reasoning ▸ Different components work in booking release different ways navigation manager Acting manager ▸ Online synthesis of automata to control their interactions … … … Nau – Lecture slides for Automated Planning and Acting 10
Example: Service Robot respond to user requests ● Multiple levels of abstraction Planning … … ▸ Higher levels: more planning bring o7 to room2 ▸ Lower levels: more acting ● Continual online planning go to navigate fetch navigate deliver ▸ What room is o7 in? hallway to room1 to room2 o7 o7 ▸ What route? … … … … ▸ What kind of door? ▸ Close enough to door handle? move to door open door get out close door ● Heterogeneous reasoning ▸ planning abstract tasks … … … ▸ path planning ungrasp ▸ reactive (e.g., open door) move identify move maintain back grasp turn Acting type close knob knob pull of pull to door knob monitor monitor Nau – Lecture slides for Automated Planning and Acting 11
Outline of Book 1: Introduction (this lecture) 2: Deterministic models ▸ Conventional ( classical ) AI planning ▸ Integrating it with acting 3: Refinement methods ▸ Acting and planning by refining abstract activities into less-abstract activities 4: Temporal models ▸ Reasoning about time constraints Automated Planning 5: Nondeterministic models ▸ Actions with multiple possible outcomes and Acting 6: Probabilistic models ▸ Multiple possible outcomes, with probabilities Malik Ghallab, Dana Nau 7: Other : and Paolo Traverso ▸ perceiving, monitoring, goal reasoning, learning, hybrid models, ontologies Nau – Lecture slides for Automated Planning and Acting 12
Any questions? Cover image: The Conjuror. Hieronymus Bosch (c.1450–1516) Nau – Lecture slides for Automated Planning and Acting 13
Recommend
More recommend
