Deliberation in Planning and Acting

Transcription

1 Last update: June 22, 2017 Deliberation in Planning and Acting Part 1: Introduction Automated Planning and Acting Malik Ghallab, Dana Nau and Paolo Traverso Malik Ghallab LAAS/CNRS, University of Toulouse Dana Nau University of Maryland Paolo Traverso FBK ICT IRST, Trento, Italy 1

2 Source Materials = Tutorial based on portions of a new book Ghallab, Nau, and Traverso (2016) Automated Planning and Acting Cambridge University Press = Free downloads at Ø Final manuscript of the book Ø Lecture slides for the book Ø Lecture slides for this tutorial Automated Planning and Acting Malik Ghallab, Dana Nau and Paolo Traverso 2

3 Motivation = Actor: agent that performs actions Ø Deliberation functions Planning What actions to perform Acting How to perform them Actor Deliberation components Queries Acting Deliberation Planning components Plans Objectives Messages Other actors = Traditional AI planning view Ø Acting = execution Ø No deliberation about how = Sometimes correct, but often incorrect Commands Percepts Execution platform Actuations Signals External World 3

4 Acting as Execution 4

5 Deliberative Acting 5

6 Planning = Relies on prediction + search s Ø Prediction uses models that predict what the actions will do Descriptive models Ø Search Search over predicted states and possible organizations of feasible actions a s = γ(s,a) = Different types of actions Ø Different predictive models Ø Different planning problems and techniques = Motion and manipulation planning = Perception planning = Navigation planning = Communication planning = Task planning 6

7 Acting = Actor is situated in a dynamic unpredictable environment Ø Adapt actions to current context Ø React to events = Relies on Ø Operational models of actions Tell how to perform the actions Ø Observations of current state Planning stage Acting stage 7

8 General Characteristics = Multiple levels of abstraction Ø Actors are organized into physical subsystems Ø Deliberation reflects this = Heterogeneous reasoning Ø Different techniques at different levels in different subsystems at same level = Continual online planning Ø Can t plan everything in advance Ø Plans are abstract and partial until more detail is needed Actor Deliberation components Objectives Deliberation Planning Queries components Plans Acting Messages Commands Percepts Execution platform Actuations Signals External World Other actors 8

9 Levels of Abstraction respond to user requests bring o7 to room2 go to hallway navigate to room1 fetch o7 navigate to room2 deliver o7 Planning move to door open door get out close door identify type of door move close to knob grasp knob turn knob maintain pull monitor move back pull monitor ungrasp Acting 9

10 Outline 1. Introduction 2. Deterministic models Ø Refinement methods for acting Ø Refinement methods for planning 3. Temporal models Ø Timelines and temporal refinement methods Ø Chronicles for planning and acting 4. Nondeterministic models Ø Offline and online nondeterministic planning Ø I/O automata and refinement methods for planning and acting 5. Conclusion Download the tutorial slides at 10

11 Relation to the Book Any questions? = Ghallab, Nau, and Traverso (2016). Automated Planning and Acting. Cambridge University Press = Free downloads: Ø Lecture slides, final manuscript Ø = Table of Contents 1. Introduction 2. Deterministic Models 3. Refinement Methods 4. Temporal Models 5. Nondeterministic Models Cover image: The Conjuror. Hieronymus Bosch (c ) 6. Probabilistic Models 7. Other Deliberation Functions 11