Multi-Layered Architectures for Autonomous Systems Jos Carlos - PowerPoint PPT Presentation

Doct. of Computer Science and Technology Ph. D. thesis defense Multi-Layered Architectures for Autonomous Systems José Carlos González Dorado Fernando Fernández Rebollo Thesis advisors Ángel García Olaya Planning and Learning Group March 27 th , 2020 Computer Science Department

Outline 1. Introduction 2. NAOTherapist architecture 3. Proposed guidelines 4. Mlaras architecture 5. Mlaras opportunities & failures 6. Mlaras in the logistics league 7. Conclusions Multi-Layered Architectures for Autonomous Systems

Motivation • Autonomous robotics is complex Rehabilitation  Many types of sensors and actuators Geriatrics  Deliberation to behave coherently Butler Logistics • Deliberation is complex Thesis  Now feasible for real systems • Architecture for coordination [Kortenkamp et al. 2008] • Use cases must be defined Logistics Simulation  How to really implement them? Multi-Layered Architectures for Autonomous Systems Introduction 1 /42 NAOTherapist

Managing use cases • Use cases can be seen as action plans [Ghallab et al. 2014] • Stochastic environments 5. Move 1. Door 6. Door 2. Move  May invalidate the plans 7. Knock 3. Make 4. Get 8. Give • Reasoning strategy  Deliberative: all deliberation (sense-plan-act), but slow  Reactive: all reactive, fast, but short-term  Hybrid : joins both, hierarchical, layered Multi-Layered Architectures for Autonomous Systems Introduction 2 /42 NAOTherapist

Control strategies • Procedural control Action decomposition ∅  Behavioral trees [Ögren et al. 2018]  Tree sets can model use cases ? Selector  Action decompositions save deliberation time • Deliberation → . . . . . . . . . . . .  Decomposed deliberation Sequential  Higher layers: deliberative → →  Lower layers: reactive → →  Several simpler problems are easier Parallel Multi-Layered Architectures for Autonomous Systems Introduction 3 /42 NAOTherapist

PELEA • Planning and execution system 8 Decision High to Low  Focused on classical planning Support AP 5  Automated Planner as a black box 6  Modular and extensible Low to High Monitoring • Ad-hoc abstraction translation 2 4 3 9 7  Actions: High to low Executive  States: Low to high 1 Low • Monitors the execution Actions Set Low State 10  Replans when state is invalid Robot [Alcázar et al. 2010] Multi-Layered Architectures for Autonomous Systems Introduction 4 /42 NAOTherapist

Layered architectures Deliberation Stochastic Temporal Declarative Multilayer Middleware ✓ LAAS 1 Cust. AP Ad hoc Cust. AP Ad hoc - ✓ T-REX 2 AP Replan Timelines Partial - ✗ PELEA 3 AP Replan Temp. AP Partial - ✗ ROSPlan 4 AP Replan Temp. AP Partial ROS ✗ CORTEX 5 Ad hoc Ad hoc Ad hoc Ad hoc RoboComp 1. [Alami et al. 1998] • Ascending order by year 2. [McGann et al. 2008] 3. [Alcázar et al. 2010] • No one fulfills everything 4. [Cashmore et al. 2015] 5. [Bustos et al. 2019] Multi-Layered Architectures for Autonomous Systems Introduction 5 /42 NAOTherapist

Problems of the architectures • Lack of guidelines and standards • Difficult to reuse previous works • Use cases are hardcoded by developers  However, they must be defined by end users • Hardcoded abstraction conversions • Complex and slow deliberation models • Lack of multilayer deliberation support All them slow down the development and advancement of autonomous systems Multi-Layered Architectures for Autonomous Systems Introduction 6 /42 NAOTherapist

Objectives of the thesis • Ease the use of standard deliberative techniques in use-case oriented cognitive architectures for autonomous systems I. Design architectures whose structures reflect the use-case  Use formalisms to involve the user in the behavior development  Ease the use-case modular decomposition into subproblems  Design layered architectures to organize knowledge II. Define relations among architecture components and layers III. Use state-of-the-art deliberative techniques IV. Carry out objectives I, II and III declaratively V. Design guidelines to apply deliberation in these systems VI. Evaluate all previous objectives in real systems Multi-Layered Architectures for Autonomous Systems Introduction 7 /42 NAOTherapist

Used deliberation techniques • Automated Planning (AP) [Ghallab et al. 2004]  Generic planner finds action plans for goals  Declarative formal language (PDDL) a b  Need to interleave planning and execution Initial state  Multiple paradigms ‒ Classical , probabilistic, temporal Planner • Mixed Integer Programming (MIP)  Fast way to reason with numbers and time a  Declarative rules and problem [Chen et al. 2010] b • More complex paradigms are slower Final state Multi-Layered Architectures for Autonomous Systems Introduction 8 /42 NAOTherapist

NAOTherapist architecture 1. Introduction 2. NAOTherapist architecture 3. Proposed guidelines 4. Mlaras architecture 5. Mlaras opportunities & failures 6. Mlaras in the logistics league 7. Conclusions Introduction Multi-Layered Architectures for Autonomous Systems NAOTherapist 9 /42 Guidelines

Mirror-game use case youtu.be/PbfqoILctH4 Introduction Multi-Layered Architectures for Autonomous Systems NAOTherapist 9 /42 Guidelines

Therapeutic protocol Therapy Sessions Exercises Poses Introduction Multi-Layered Architectures for Autonomous Systems NAOTherapist 10 /42 Guidelines

Use-case planning levels • High level: plan all therapy sessions  Classical planning  Many complex constraints Therapy  Offline, manual replannings Sessions • Medium level: plan the actual execution Exercises  Classical planning and PELEA to replan online Poses  Problems converted from the high level  Centralized ad-hoc abstraction translations • Low level: plan each movement  Transparent and independent from the robot and sensor Introduction Multi-Layered Architectures for Autonomous Systems NAOTherapist 11 /42 Guidelines

High-level therapy designer Introduction Multi-Layered Architectures for Autonomous Systems NAOTherapist 12 /42 Guidelines

Medium and low-level architecture Introduction Multi-Layered Architectures for Autonomous Systems NAOTherapist 13 /42 Guidelines

Technical evaluations • Planning times  62 real short sessions  23 poses each • Generalization  Simon  Reverse Simon  Simon says [García et al. 2017]  Dancing  Teaching movements Introduction Multi-Layered Architectures for Autonomous Systems NAOTherapist 14 /42 Guidelines

Field tests • 3 field tests with patients  3 children, 1 session  8 children, 4 months, weekly  10 children, 15 days, daily • Overall  244 children (21 patients)  429 sessions (206 of patients)  Good interactive outcomes • Clinical and HRI analysis [Pulido 2020] Introduction Multi-Layered Architectures for Autonomous Systems NAOTherapist 15 /42 Guidelines

Lessons learned • Multilayered deliberation works well • Needs guidelines • Need to fix the monolithic Executive ✓ Rehabilitation  Hard to maintain, small changes hard to apply ✗ Geriatrics • Needs a full declarative configuration ✗ Butler ✗ • Needs action interruption Logistics • Needs an online high-level planning • The NAOTherapist architecture is not enough Introduction Multi-Layered Architectures for Autonomous Systems NAOTherapist 16 /42 Guidelines

Proposed guidelines 1. Introduction 2. NAOTherapist architecture 3. Proposed guidelines 4. Mlaras architecture 5. Mlaras opportunities & failures 6. Mlaras in the logistics league 7. Conclusions NAOTherapist Multi-Layered Architectures for Autonomous Systems Guidelines 17 /42 Mlaras architecture

Objective of the guidelines • Face the system design from the use case 1. Deliberation strategy: problem decomposition 2. Planning model: domain and problem modeling 3. Executive model: interleaving planning and execution 1. Design deliberation strategy 2. Design planning model 3. Design executive model NAOTherapist Multi-Layered Architectures for Autonomous Systems Guidelines 17 /42 Mlaras architecture

Guidelines for AP and robotics 1. Design deliberation strategy  Define planning paradigm ‒ Classical, temporal, probabilistic…  Design layer reasoning ‒ Deliberative, reactor Layered decomposition P A A A 1 A A 2 A A 3 A A 4 Problem P B A B 1 A B 2 A B 3 A B 4 P A Define therapy A A Session definition NAOTherapist P B Execute session A B Interactive actions NAOTherapist Multi-Layered Architectures for Autonomous Systems Guidelines 18 /42 Mlaras architecture

Guidelines for AP and robotics 1. Design deliberation strategy  Design planning decomposition Therapy High-level Planning horizon Divide & conquer P A A A 1 A A 2 A A 3 A A 4 Sessions Problem Exercises P B1 A B1 1 A B1 2 P B2 A B2 1 Poses P B1 Session 1 NAOTherapist P B2 Session 2 A B Exercise NAOTherapist Multi-Layered Architectures for Autonomous Systems Guidelines 19 /42 Mlaras architecture