Timeline-based Planning and Execution: Theory and Practice

- PLATINUm -A Novel framework for PLanning and Acting

with TImeliNes under Uncertainty

Alessandro Umbrico National Research Council of Italy (ISTC-CNR)

Outlinep General Introduction to PLATINUm

n History, general motivations and objectivesp PLATINUm Representation Capabilities

n Temporal Uncertainty, Components and Resolversp PLATINUm Deliberative Capabilities

n Pseudo-controllability aware planning and hierarchical approachp PLATINUm Executive Capabilities

n Closed-loop control approach and controllability issuesp PLATINUm in Action

n Human-Robot Collaboration in realistic manufacturing scenarios1

GENERAL INTRODUCTION TO PLATINUM

2

A Brief History of PLATINUm

3

HSTS[Muscettola et al. 1994]

APSI-TRF[Cesta et al. 2008]

APSI-GOAC[Fratini et al. 2011]

EPSL[Umbrico et al. 2015]

PLATINUm[Umbrico et al. 2017]

Formalization of Timeline-based Approach with Temporal Uncertainty and Resources

[Cialdea et al. 2016] + [Umbrico et al. 2018]

PLATINUm+ Resources

[Umbrico et al. 2018]

Motivations and Objectivesp Several timeline-based systems successfully applied in practice

n E.g., EUROPA2, ASPEN, IxTeT, APSI-TRF

p Lack of uniform formal interpretation of the main concepts n E.g., different interpretation of timelines, domain rules, plans and solutions

p Lack of a uniform approach to planning and execution with timelinesn E.g., different approaches to solving and modeling problems with timelines

p Lack of representation of uncertainty and uncontrollable dynamics

4

Framework Capabilitiesp A software framework capable of dealing with planning and

execution of timelines under temporal uncertaintyn Comply with the formalization proposed in [Cialdea et al. 2016]

p Synthesize timeline-based plans with some desired controllability propertiesn E.g., pseudo-controllability

p Execute timelines by dealing with the dynamics of the uncontrollable features of the environmentn The controllability problem [Morris, Muscettola and Vidal 2001]

5

Architectural Overview of the Framework

Deliberative

Planner

Strategy Solver Heuristics

Representation

PlanDatabase

DomainComponent

Resolver

TemporalDatabase ParameterDatabase

Executive

Executive

Monitor Clock Dispatcher

ExecutivePlanDatabaseManager

6

A Layered Architecture for PLATINUmp The Representation Layer is responsible for providing the basic

functionalities needed to manage timeline-based plansn Encapsulate data structures and algorithms needed to build valid timelines

p The Deliberative Layer is responsible for solving planning problems by synthesizing timeline-based solution plansn Encapsulate heuristics, strategies and algorithms needed to build plans

p The Executive Layer is responsible for executing timelines by scheduling flexible tokens over timen Encapsulate dispatching and monitoring algorithms to dynamically adapt

timelines during execution, according to the received feedbacks 7

GOING DEEP INTO REPRESENTATION CAPABILITIES

8

Plan Database Overall Structure

9

Plan Data Representation and Managementp Temporal information is managed through Temporal Networks

n Framework enabling temporal inference and consistency checking [Dechter et al. 1991]

p Simple Temporal Network with Uncertainty (STNU) to manage uncontrollable durations

p Controllability check [Morris, Muscettola and Vidal 2001]

p A “standard” CSP solver is encapsulated to manage variable assignment and constraint propagationn Choco CSP solver

10

Resolvers and Componentsp Domain components are data structures that model the features

of a planning domain that must be controlled over timen Each feature has its own constraints that must be satisfied to generate

valid timelines i.e., temporal behaviors without flaws

p Resolvers represent specialized algorithms capable of detecting flaws and computing possible solutionsn Each resolver encapsulates the logic for handling a particular type of flaw

p PLATINUm provides a set of ready to use components and resolvers representing the typical features that compose a timeline-based planning domain

11

State Variablesp Components that comply with

the proposed formalizationn Encapsulate values, durations,

controllability tags and transition constraints

p Resolvers are provided to synthesize timelinesn Schedule state variable tokensn Synthesize complete sequences

of tokens enforcing transition constraints

12

Discrete and Reservoir Resourcesp Components encapsulate

resource constraints and resource events according to the tokens of the timelines

p Resolvers provide the logic for detecting and solving peaksn Compute pessimistic and

optimistic resource profilesn Compute peak solutions through

Minimal Critical Sets (MCSs)p Planning & Scheduling integration 13

The Plan Databasep Encapsulating all the domain

components and configurationsn Composite design patternn Provide a public interface to

domain features and data

p Leverage internal components to detect planning goalsn Planning goal solutions computed

through synchronization rulesn Apply expansion or unification

14

GOING DEEP INTO DELIBERATIVE CAPABILITIES

15

Detailed Structure of a PLATINUm Planner

16

Pseudo-controllability Aware Solvingp General plan refinement search procedure

n Iteratively refine an initial partial plan by solving flaws, until a valid plan without flaws is found

n Search decision point: which partial plan to select for search space expansion

n Flaw decision point: which flaw to solve for plan refinement

p Pseudo-controllability check as a special flaw of the plann Verify if the flexible durations of the uncontrollable tokens have been

modified with respect to the domain specificationn Pseudo-controllability is a necessary but not sufficient condition for

dynamic controllability 17

Hierarchical Flaw Selection Heuristicsp Analyze synchronization rules

to extract dependencies among state variables and their valuesn Extract a hierarchy if possiblen Domain-independent heuristics

p Leverage the extracted hierarchy to evaluate flaws and decide which flaw to solve nextn Support flaw decision point

18

Functional variables

SVA

SVB SVC

Primitivevariables

SVD SVE

SVA

SVB

SVC

SVD

SVE

Search & Build Plan Synthesisp Search phase aims at constraining temporal behaviors as much

as possiblen Interleave planning and scheduling decisions by reasoning on flawsn Constraint state variables behaviors according to synchronization rules

and resource constraints - generated behaviors are not timelines yet

p Build phase aims at finalizing timelines by enforcing semantics defined by the formalizationn Synthesize valid timelines according to the constrained behaviors of

state variables generated by the search phasen Backtrack by jumping-back to the search phase in case of failures

19

Defining and Using a PLATINUm Planner

20

GOING DEEP INTO EXECUTIVE CAPABILITIES

21

Detailed Structure of a PLATINUm Executive

22

Closed-loop Execution of Timeline-based Plansp A PLATINUm executive consists of a closed-loop control

process which iteratively fix flexible timelines over time

p A Dispatcher actually executes the timelines of a plan by sending commands to a physical systemn It is responsible for deciding the start of the execution of the tokens that

compose the timelines of a plan

p A Monitor handles execution feedbacks to verify whether the plan complies with the observed status of the environment or notn It is responsible for propagating information about the actual duration of

uncontrollable tokens23

Executing Timelines under Temporal Uncertaintyp Extract start/end execution

dependencies by analyzing temporal relations of a plann Dynamically generate a

execution dependency graph

p Manage token transitions according to their controllability propertiesn Different controllability properties

entail different dispatching policies

waiting

in-execution

c

executed

c

controllable

waiting

in-execution

c

executed

u

partially controllable waiting

starting

c

in-execution

u

executed

u

uncontrollable

24

Closed-loop Control Architecture

Deliberativebuffered

planned

FailureManager

Executive

Dispatcher Monitor

failureexecuted

re-planning

System/ROS-basedSimulator

send command feedback feedbacksend command

25

HUMAN-ROBOT COLLABORATION CASE STUDY

26

The FourByThree Research Projectp Horizon 2020 research project

n Call FoF-06-2014 “Symbiotic Human-Robot Collaboration for safe and dynamic multimodal manufacturing systems”

n Coordinated by FUNDACION TEKNIKER (Spain)n http://fourbythree.eu/

27

Objectives of the Projectp Develop a new generation of modular industrial robotic solutions

that are suitable for efficient task execution in collaboration with humans in a safe way and are easy to use and programn Vision: A system integrator (or end-user) can create its own custom robot

according to their application needs (“kit” of hardware and software components)

p Real-world Pilot case studies to test Human-Robot Collaborationn ALFA, WOLL, STODT, PREMIUM

28

ALFA: A Collaborative Assembly Case Studyp Investment Casting process

n Dies are assembled and disassembled manually

n Some operations need human dexterityn Others can be done by a robot

p Re-design the process by taking into account an HRC perspectiven Hierarchical process descriptionn Three levels: Supervision, Coordination,

Implementation29

A Hierarchical Planning Model for Collaborative Assembly Scenarios

Robot’smotiontasks

Human’sassignedtasks

30

Flexible Collaboration with Dynamic Trajectory Selection

Verify the capabilities of an integrated task and motion planning control to improve the efficiency of HRC assembly processes

31

Experimental Results

257

225

195171 169 172

0%

10%

20%

30%

40%

50%

60%

70%

80%

50

100

150

200

250

300

0 1 2 3 4 5 6 7

ExecutionTIme[s]

ExecutionTime[s] SavedTime[%] RobotAllocatedTask[%]

Planning can reduce the duration of HRC processes and realize effective and safe collaborations between human operators and

robots [Pellegrinelli et al. CIRP Annals 2017] 32

Final Remarksp HRC is particularly suited to show the capabilities of PLATINUm

p Temporal uncertainty and hierarchical problem modeling and solving

p The FourByThree research project has shown the capability of PLATINUm to synthesize flexible and adaptable behaviors of a robot acting in the real-world

p Download PLATINUm from GitHub:https://github.com/pstlab/PLATINUm

33

CORRELATED PROJECTS AND FUTURE DEVELOPMENTS

34

KEEN– The Knowledge Engineering Environment for Timeline-based Planningp A knowledge engineering environment aimed at supporting the

design and development of timeline-based applicationsn Round-trip-engineering to support both graphical and programming user

interface to model timeline-based domainsn Available as plugin for the Eclipse IDEn https://ugilio.github.io/keen/

p Integrated with PLATINUm n Launch platinum-based planners on designed timeline-based domainsn Launch platinum-based executive to simulate execution of synthesized

plans35

Conclusions and Future Worksp PLATINUm is an open framework which can be “easily” extended

to introduce new representation features and solving capabilitiesn Discrete and Reservoir Resource management has been recently

introduced [Umbrico et al. ICAPS 2018]

p Future worksn Investigate new heuristics and new search strategiesn Integrate dynamic controllability checking algorithms n Comparison with other state of the art timeline-based and hybrid planners

p EUROPA [Barreiro et al., 2012], CHIMP [Stock et al., 2015], FAPE [Dvorak et al., 2014]

36

Thanks for your Attention!

37

PLATINUm – PLanning and Actingwith TImeliNes under Uncertainty

Available on GitHubhttps://github.com/pstlab/PLATINUm