Artificial Intelligence: Methods and Applications Lecture 3: Hybrid - PDF document

Artificial Intelligence: Methods and Applications Lecture 3: Hybrid robot architechtures Henrik Björklund Umeå University 27. November 2012 Thesis: The deliberative paradigm ◮ ca 1967 - ca 1990 ◮ AI inspired ◮ Represent every relevant aspect of the world explicitly ◮ Interpret sensor data: make it a part of the world model ◮ Use classical planning to decide what to do

Deliberative paradigm Pro: ◮ Goal oriented ◮ Predictable ◮ Clear and sound reasoning Con: ◮ Computationally expensive ◮ Frame problem: actions can have many effects ◮ Requires exact knowledge of the world ◮ Symbol grounding problem ◮ Discretization Antithesis: The reactive paradigm ◮ ca 1988 - ca 1992 ◮ A reaction to classical AI ◮ Less knowledge representation and planning ◮ More concrete repsonses to the environment ◮ Decompose complex actions into behaviors

Reactive paradigm Pro: ◮ Short reaction times ◮ Needs less computational resources ◮ Easy to implement and expand ◮ Emergent behavior ◮ Open world assumption Con: ◮ Unpredictable ◮ Unclear reasoning ◮ No monitoring of performance ◮ No world representation (no internal map) ◮ No selection of behaviors ◮ More art than science? Synthesis: The hybrid reactive/deliberative paradigm Slogan: The best of both worlds! Caveat: See to it that that is actually what you get! ◮ How to reintroduce planning into the robot architectures without running into the problems that faced deliberative robots? ◮ Use reactive functions for low level control ◮ Use deliberation for higher level tasks

Combining deliberative and reactive functions ◮ Deliberative: ◮ Long time horizon ◮ Global knowledge ◮ Works with symbols ◮ Reaction: ◮ Short time horizon ◮ No global knowledge ◮ Works with sensors and actuators ◮ Multi-tasking: ◮ Deliberative and reactive functions execute in parallel Deliberative: Sense, Plan, Act Sense Plan Act

Reactive: Sense and Act Sense Act Hybrid: Plan || Sense and Act Plan Sense Act

What should the planning component do? ◮ Manage behaviors ◮ What is the current state of the world? ◮ What is the goal state? ◮ Which (combinations of / sequences of) behaviors will achieve the goal? ◮ Monitor performance ◮ Was the latest sub-plan successful? ◮ Are sensors and actuators working properly? ◮ Is the sensor data compatible with my view of the world? ◮ If sensors are giving contradictory data, what to do? Common components Most hybrid architectrures incorporate (variants of) the following components: ◮ Mission planner ◮ Interpret commands and create a high-level plan ◮ Sequencer ◮ Given a sub-task, generate a sequence of behaviors to solve it ◮ Resource manager ◮ Allocate recources to behaviors ◮ Performance monitor ◮ Determine if the robot is functioning properly and making progress towards the goals ◮ Cartographer ◮ Create, store, and maintain spatial data

Autonomous Robot Architecture (AuRA) ◮ Suggested in the mid 1980s ◮ Ronald C. Arkin ◮ First hybrid architecture ◮ Georgia Tech ◮ Based on schema theory ◮ Nested Hierarchical Controller ◮ Potential fields for motor schemas AuRA Structure

AuRA Structure Planner Mission planner Navigator Pilot AuRA Structure Planner Mission planner Navigator Cartographer Pilot

AuRA Structure Planner Mission planner Navigator Cartographer Pilot Sensors PS1 PS2 PS3 AuRA Structure Planner Mission planner Navigator Cartographer Pilot Sensors Motor schema manager PS1 MS1 PS2 MS2 Σ PS3 MS3

AuRA Structure Planner Mission planner Navigator Cartographer Homeostatic Control Pilot Sensors Motor schema manager PS1 MS1 PS2 MS2 Σ PS3 MS3 AuRA Summary Mission planner Mission planner Sequencer Navigator, Pilot Recource manager Motor schema manager Performance monitor Pilot, Navigator, Mission planner Cartographer Cartographer Emergent Vector sums, spreading activation, homeostatic control

State-Hierarchy Architectures Activities are organized by time scope or state of knowledge. Usually, there are three layers: 1. Future 2. Past 3. Present 3T 3T is a state-hierarchy architecture that has been extensively used by NASA. ◮ Ca 1996 ◮ Merges subsumption (Gat, Bonasso), RAPs (Firby), and vision (Kortenkamp) ◮ Three layers: ◮ Deliberative ◮ In-between (reactive planning) ◮ Reactive ◮ Arranges tasks by execution rate ◮ Planetary rovers ◮ Underwater vehicles

3T Structure 3T Structure Goals Planner

3T Structure Goals Planner Arrange tasks Task commitments Sequencer 3T Structure Goals Planner Arrange tasks Task commitments Sequencer Configure Signals Skill manager Actuators Sensors

3T Summary Mission planner Planner Sequencer Sequencer Recource manager Sequencer Performance monitor Planner Cartographer Planner Emergent Behaviours grouped into skills, skills grouped into task networks Model-oriented architectures Where managerial or state-hierarchy architectures have a bottom-up flavor, model-oriented architectures are more top-down. ◮ Based around a global world model ◮ More focus on classical AI and (somewhat) less on biologically inspired reactive features ◮ Sensor data filtered through the global world model ◮ Less ambitious world model ◮ Distributed processing of sensor data ◮ Assign symbolic labels to map items

Saphira ◮ Ca 1997-onwards ◮ Kurt Konolige et al. ◮ SRI International ◮ Flakey, Erratic Motivation: ◮ Coordination ◮ Coherence ◮ Communication Saphira structure

Saphira structure Local Perceptual Space Saphira structure PRS-lite (planning) Local Perceptual Space

Saphira structure PRS-lite (planning) Tracking Recognition Local Perceptual Surfaces Space Localization Saphira structure PRS-lite (planning) Tracking Recognition Local Perceptual Surfaces Space Localization Sensors

Saphira structure PRS-lite (planning) Tracking Recognition Top. planner Local Perceptual Surfaces Space Navigation Localization Sensors Saphira structure PRS-lite (planning) Tracking Recognition Top. planner Local Perceptual Surfaces Space Navigation Localization Sensors Behaviours

Saphira structure PRS-lite (planning) Tracking Recognition Top. planner Local Perceptual Surfaces Space Navigation Localization Fuzzy Sensors Behaviours logic Saphira Summary Mission planner PRS-lite Sequencer Topological planner, Navigation Recource manager PRS-lite Performance monitor PRS-lite Cartographer LPS Emergent Behaviours fused with fuzzy logic

Task Control Architecture (TCA) ◮ Reid Simmons ◮ Used by NASA robots ◮ CMU Xavier ◮ Doesn’t use behaviours DARPA Grand Challenge The 2007 Urban Challenge, 96 km in an urban environment, was won by Tartan Racing from Carnegie Mellon University.

Deliberative vs. Hybrid Do deliberative and hybrid architectures simply come to the same conclusions in different ways? ◮ Hybrids are closer to software engeneering principles ◮ In hybrid architectures, the world model is only used on a high level ◮ Use symbolic representation for high-level ”thinking” ◮ The frame problem is not much of a problem for hybrids ◮ Think in terms of a closed world ◮ Act and sense in an open world ◮ Deliberative functions in hybrid architectures don’t have to do detailed planning ◮ Hybrids can be relevant for cognitive science