Task Monitoring and Rescheduling for Opportunity and Failure - - PowerPoint PPT Presentation

▶
task monitoring and rescheduling for opportunity and
SMART_READER_LITE
LIVE PREVIEW

Task Monitoring and Rescheduling for Opportunity and Failure - - PowerPoint PPT Presentation

Oct 19, 2022 313 likes •567 views

IntEx Workshop Task Monitoring and Rescheduling for Opportunity and Failure Management Jos Carlos Gonzlez, Manuela Veloso, Fernando Fernndez and ngel Garca-Olaya Planning and Learning Group 25 June 2018 Computer Science Department

slide-1
SLIDE 1

Task Monitoring and Rescheduling for Opportunity and Failure Management

José Carlos González, Manuela Veloso, Fernando Fernández and Ángel García-Olaya

25 June 2018 Computer Science Department

Planning and Learning Group

IntEx Workshop

slide-2
SLIDE 2

2/20

Introduction – Tasks of service robots

  • Robot must find a valid task schedule, and execute it
  • Several constraints per task
  • Users can add tasks anytime

Task Monitoring and Rescheduling for Opportunity and Failure Management

Go to a place Deliver message Escort someone Deliver object Make Coffee Bring message Remind something Recharge battery

Users Pending task pool Robot

Introduction

Opportunities and Failures

slide-3
SLIDE 3

3/20

Introduction – Hot coffee delivering

Task Monitoring and Rescheduling for Opportunity and Failure Management

A B

Introduction

Opportunities and Failures

Subtasks: A, B

slide-4
SLIDE 4

4/20

Task Monitoring and Rescheduling for Opportunity and Failure Management

B

Quick, or it’ll get cold!

Opportunity (finish the task earlier) Failure (nobody is in the office)

Introduction – Hot coffee delivering

Introduction

Opportunities and Failures

Subtasks: A, B

slide-5
SLIDE 5

5/20 A B

VIP

Task Monitoring and Rescheduling for Opportunity and Failure Management

Introduction – Hot coffee delivering

Introduction

Opportunities and Failures

Subtasks: A, B

slide-6
SLIDE 6

6/20 B

VIP !

Task Monitoring and Rescheduling for Opportunity and Failure Management

Opportunity (high-priority task)

Introduction – Hot coffee delivering

Cooling-down time

Introduction

Opportunities and Failures

Subtasks: A, B

slide-7
SLIDE 7

7/20 B

VIP !

Task Monitoring and Rescheduling for Opportunity and Failure Management

What to do now?

  • VIP first, then resume B
  • Redo A and B
  • VIP after B
  • Cancel A and B
  • Cancel VIP
  • Try a quick VIP

A

Introduction – Hot coffee delivering

Introduction

Opportunities and Failures

slide-8
SLIDE 8

8/20

Opportunities and Failures

Task Monitoring and Rescheduling for Opportunity and Failure Management

Introduction

Opportunities and Failures

Modeling

Opportunities: Failures:

Current task

Opportunities: Failures:

Next task

. . .

Scheduler

Min total time Max total priority

State

Constraints Priority: 5 Constraints Priority: 1

slide-9
SLIDE 9

9/20

Opportunities and Failures

Task Monitoring and Rescheduling for Opportunity and Failure Management

Introduction

Opportunities and Failures

Modeling

Opportunities: Failures:

Current task

Opportunities: Failures:

Next task

. . .

State

Constraints Priority: 5 Constraints Priority: 1

High-level events must be checked for all scheduled tasks

Scheduler

Min total time Max total priority

Reschedule!

slide-10
SLIDE 10

10/20

Contribution and Related work

  • Our contribution

▪ Component to handle high-level unexpected events among tasks ▪ MIP model with dependent tasks and cooling-down times

  • Coltin, B.; Veloso, M. M.; and Ventura, R. 2011.

Dynamic user task scheduling for mobile robots

▪ Fixed schedules with a Mixed Integer Programming (MIP) solver

  • Cashmore, M.; Fox, M.; Long, D.; et al. 2017.

Opportunistic Planning in Autonomous Underwater Missions

  • Schermerhorn, P.; Benton, J.; Scheutz, M.; et al. 2009.

Finding and Exploiting Goal Opportunities in Real-Time During Plan Execution

Task Monitoring and Rescheduling for Opportunity and Failure Management

Our starting point

Introduction

Opportunities and Failures

Modeling

slide-11
SLIDE 11

11/20

Monitoring model

  • Updated states received while

subtasks are being executed

  • Generic task attributes

Opportunities and Failures

▪ Indicate parameters in the state that should remain invariant ▪ Used to trigger reschedulings

  • A rescheduling can

▪ Add or remove tasks in the pool ▪ Interrupt the current subtask

Task Monitoring and Rescheduling for Opportunity and Failure Management

Lower Levels High-level Scheduler Task Pool

Task 1 Task n

Opportunities Failures

. . .

State Tasks

Opportunities and Failures

Modeling

Experiments

slide-12
SLIDE 12

12/20

High-level Task Scheduler Architecture

Task pool Problem Schedule

Knowl. Base Execution Monitoring

Opportunities Failures Interruptions

Solver

  • Ext. MIP solver

Tasks Data State Task Task Static data

Scheduler User Interface

Tasks

Robot

State Tasks Task Monitoring and Rescheduling for Opportunity and Failure Management

  • Multilevel global scheme

▪ Rescheduling for high-level events ▪ Tasks sent to lower abstraction levels ▪ States are generalized from lower levels

Opportunities and Failures

Modeling

Experiments

slide-13
SLIDE 13

13/20

Task modeling and decomposition

Task Subtask-1 Subtask-2

Task type Task owner Location start Location end Time start min Time end max Person target Object Priority DeliverDrink Alice

  • 15

Alice HotCoffee

  • MakeHotDrink

Alice CoffeMaker CoffeMaker 15

  • HotCoffee

10 DeliverObject Alice CoffeMaker AliceOffice 15 Alice HotCoffee 10 Time operation Time cooldown Task depending Opportunities Failures

  • VIP

TO, BP 5

  • HotCoffee, VIP

TO, BP 2 6

Subtask-1

Person target, VIP HotCoffee, TO, BP

User Internal

Task Monitoring and Rescheduling for Opportunity and Failure Management

Opportunities and Failures

Modeling

Experiments

slide-14
SLIDE 14

14/20

MIP model with cooling-down time

Task Monitoring and Rescheduling for Opportunity and Failure Management

Depending subtasks and cooling-down Order and overlapping

Solution types

▪ Proven optimal ▪ Suboptimal ▪ Not found

‒ Unfeasible ‒ Time limit Opportunities and Failures

Modeling

Experiments

slide-15
SLIDE 15

15/20

Rescheduling policy

  • If the scheduler cannot find a suitable plan

▪ Failures: Monitoring cancels the next task

‒ With the lowest priority first ‒ Then the smallest time window that overlaps another

▪ Opportunities:

  • 1. Tries to redo the current subtask later
  • 2. If it cannot, it tries to redo the whole task
  • 3. If it cannot, it evaluates whether to cancel the current task
  • r the new task by maximizing the gain measure g

Task Monitoring and Rescheduling for Opportunity and Failure Management

Sum of the priorities of the scheduled tasks

Opportunities and Failures

Modeling

Experiments

slide-16
SLIDE 16

16/20

Experiments – CoBot robots

Task Monitoring and Rescheduling for Opportunity and Failure Management

  • Using the CoBot platform
  • Their task catalog
  • Schedules work in the

actual robot

  • 180 simulations
  • Scheduling times
  • Quality

Modeling

Experiments

Conclusions

slide-17
SLIDE 17

17/20

Experiments – Schedules

Schedule 2 Schedule 3 Schedule 1

Task Start End … 10 C1a 11 20 C2a 21 26 C1b 27 31 C2b 32 33 VIP 34 39 C3a 40 45 C3b 46 50 Cost 605 Task Start End … 10 C1a 11 20 C2a 21 26 C1b 27 31 C2b 32 33 C3a 34 42 C3b 43 47 VIP 48 53 Cost 739 Task Start End … 10 C1a 11 20 VIP 21 23 C2a 24 29 C1b 30 34 C2b 35 36 C3a 37 45 C3b 46 50 Cost 454

Task Monitoring and Rescheduling for Opportunity and Failure Management

  • Task decomposition allows to optimize locations

Modeling

Experiments

Conclusions

slide-18
SLIDE 18

18/20

Experiments – Solving time vs. Subtasks

5 10 15 20 25 30 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

  • Avg. solving time (s)

Task pool size (set B)

10 s, 4.4% tol. 10 s 30 s

Task Monitoring and Rescheduling for Opportunity and Failure Management

(all with solutions)

  • Proven optimal solutions found up to size 10

(suboptimal)

Modeling

Experiments

Conclusions

slide-19
SLIDE 19

19/20

Experiments – Quality vs. Subtasks

500 550 600 650 700 750 800 850 0 1 2 3 4 5 6 7 8 9 101112131415

  • Avg. quality (min)

Task pool size (set B)

10 s, 4.4% tol. 10 s 30 s

Task Monitoring and Rescheduling for Opportunity and Failure Management

(all with solutions)

(suboptimal)

  • Quality in “10s suboptimal” is acceptable for the CoBot’s domain

Modeling

Experiments

Conclusions

slide-20
SLIDE 20

20/20

Conclusions

  • New architecture of task execution, monitoring and rescheduling

▪ Rescheduling according to opportunities and failures ▪ Interruption of tasks in the middle of their execution ▪ Future work: integration with a generic hierarchical control architecture, independent from the planning/scheduling mechanism

  • Improved MIP model

▪ Able to deal with cooling-down times and dependent tasks ▪ Focused on the quality of the solutions ▪ Quality can be affected in extreme conditions with large task pools and fast solving times required ▪ Future work:

‒ Transform some hard-constraints (time-window) into soft ‒ Comparisons with other rescheduling systems

Task Monitoring and Rescheduling for Opportunity and Failure Management

Experiments

Conclusions

slide-21
SLIDE 21

Task Monitoring and Rescheduling for Opportunity and Failure Management

José Carlos González, Manuela Veloso, Fernando Fernández and Ángel García-Olaya Planning and Learning Group

Thank you for your attention

25 June 2018 Computer Science Department

IntEx Workshop

slide-22
SLIDE 22

22/20

Opportunities and Failures

  • High-level events

▪ Affect the current task and future tasks in the schedule ▪ Interrupt tasks in the middle of their execution

  • Opportunities

▪ Domain: can appear at any moment (VIP) ▪ Specific: exclusive for a task (receipt of the coffee found earlier)

  • Failures

▪ Domain: same failure for several tasks (blocked paths, timeout) ▪ Specific: exclusive for a task (coffee stolen)

Task Monitoring and Rescheduling for Opportunity and Failure Management

slide-23
SLIDE 23

23/20

Experimental sets

Task Monitoring and Rescheduling for Opportunity and Failure Management

Experimental sets A>B>C

  • A: 480 random instances

(task pools)

  • B: 12 solved instances

per each pool size from 1-15 (180 in total)

  • C: 12 random instances

per each pool size from 8-15 (96 in total)

slide-24
SLIDE 24

24/20

Experiments – Solution types

Configuration 10 s, 4.4% tol.

10 s 30 s

Time out: no solut.

11.0% 11.0% 8.8%

Proven unfeasible

0.8% 0.8% 0.8%

Check failed

4.4% 4.4% 4.4%

Proven optimal

16.3% 42.7% 43.1%

  • Min. gap reached

54.0% 0.0% 0.0%

Time out: found

13.5% 41.0% 42.9%

Solutions found

83.8% 83.8% 86.0%

Proven optimal

17.8% 51.1% 52.2%

  • Min. gap reached

68.3% 0.0% 0.0%

Time out: found

13.9% 48.9% 47.8%

  • Av. solver time (s)

2.14 ± 3.6 5.07 ± 4.9 14.7 ± 14.8

  • Av. quality (min)

611 ± 256 596 ± 250 590 ± 247

Proven optimal

0.0% 10.4% 12.5%

  • Min. gap reached

74.0% 0.0% 0.0%

Time out: found

26.0% 89.6% 87.5%

  • Av. solver time (s)

3.98 ± 4.1 9.24 ± 2.5 26.88 ± 8.6

  • Av. quality (min)

738 ± 135 721 ± 137 709 ± 137

Task Monitoring and Rescheduling for Opportunity and Failure Management

Set A Set B Set C