IntEx Workshop Task Monitoring and Rescheduling for Opportunity and Failure Management José Carlos González, Manuela Veloso, Fernando Fernández and Ángel García-Olaya Planning and Learning Group 25 June 2018 Computer Science Department
Introduction – Tasks of service robots • Robot must find a valid task schedule, and execute it • Several constraints per task • Users can add tasks anytime Go to a place Deliver message Escort someone Deliver object Make Coffee Bring message Remind something Recharge battery Users Pending task pool Robot Task Monitoring and Rescheduling for Introduction 2 /20 Opportunity and Failure Management Opportunities and Failures
Introduction – Hot coffee delivering A Subtasks: A, B B Task Monitoring and Rescheduling for Introduction 3 /20 Opportunity and Failure Management Opportunities and Failures
Introduction – Hot coffee delivering Quick, or it’ll get cold! Subtasks: A, B B Opportunity (finish the task earlier) Failure (nobody is in the office) Task Monitoring and Rescheduling for Introduction 4 /20 Opportunity and Failure Management Opportunities and Failures
Introduction – Hot coffee delivering A Subtasks: A, B B VIP Task Monitoring and Rescheduling for Introduction 5 /20 Opportunity and Failure Management Opportunities and Failures
Introduction – Hot coffee delivering Subtasks: A, B B ! Cooling-down time VIP Opportunity (high-priority task) Task Monitoring and Rescheduling for Introduction 6 /20 Opportunity and Failure Management Opportunities and Failures
Introduction – Hot coffee delivering A What to do now? B • VIP first, then resume B ! • Redo A and B • VIP after B • Cancel A and B • Cancel VIP • Try a quick VIP VIP Task Monitoring and Rescheduling for Introduction 7 /20 Opportunity and Failure Management Opportunities and Failures
Opportunities and Failures Current task Opportunities: Failures: Scheduler Constraints Min total time Priority: 5 Max total priority Next task Opportunities: State Failures: Constraints Priority: 1 . . . Introduction Task Monitoring and Rescheduling for Opportunities and Failures 8 /20 Opportunity and Failure Management Modeling
Opportunities and Failures Current task High-level events must be checked for all scheduled tasks Opportunities: Failures: Scheduler Constraints Min total time Priority: 5 Max total priority Reschedule ! Next task Opportunities: State Failures: Constraints Priority: 1 . . . Introduction Task Monitoring and Rescheduling for Opportunities and Failures 9 /20 Opportunity and Failure Management Modeling
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. Our starting Dynamic user task scheduling for mobile robots point ▪ 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 Introduction Task Monitoring and Rescheduling for Opportunities and Failures 10 /20 Opportunity and Failure Management Modeling
Monitoring model Task Pool • Updated states received while Task 1 subtasks are being executed . . . • Generic task attributes Task n Opportunities and Failures Opportunities ▪ Indicate parameters in the state Failures that should remain invariant ▪ Used to trigger reschedulings High-level Scheduler • A rescheduling can ▪ Add or remove tasks in the pool State Tasks ▪ Interrupt the current subtask Lower Levels Opportunities and Failures Task Monitoring and Rescheduling for Modeling 11 /20 Opportunity and Failure Management Experiments
High-level Task Scheduler Architecture User Tasks Interface Scheduler Tasks Task • Multilevel global scheme pool Monitoring ▪ Rescheduling for high-level Opportunities Problem Failures Tasks events Interruptions ▪ Tasks sent to lower Task State Schedule abstraction levels Solver ▪ States are generalized Execution Ext. MIP solver Static from lower levels data Task State Knowl. Data Robot Base Opportunities and Failures Task Monitoring and Rescheduling for Modeling 12 /20 Opportunity and Failure Management Experiments
Task modeling and decomposition Task Subtask-1 Subtask-2 Task type DeliverDrink MakeHotDrink DeliverObject Task owner Alice Alice Alice Location start - CoffeMaker CoffeMaker Location end - CoffeMaker AliceOffice User Time start min 0 0 0 Time end max 15 15 15 Person target Alice - Alice Object HotCoffee HotCoffee HotCoffee Priority - 10 10 - 2 Time operation 5 - 6 Time cooldown - Internal - Subtask-1 Task depending - VIP Opportunities HotCoffee, VIP Person target, VIP TO, BP Failures TO, BP HotCoffee, TO, BP Opportunities and Failures Task Monitoring and Rescheduling for Modeling 13 /20 Opportunity and Failure Management Experiments
MIP model with cooling-down time Order and overlapping Depending subtasks and cooling-down Solution types ▪ Proven optimal ▪ Suboptimal ▪ Not found ‒ Unfeasible ‒ Time limit Opportunities and Failures Task Monitoring and Rescheduling for Modeling 14 /20 Opportunity and Failure Management Experiments
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 or the new task by maximizing the gain measure g Sum of the priorities of the scheduled tasks Opportunities and Failures Task Monitoring and Rescheduling for Modeling 15 /20 Opportunity and Failure Management Experiments
Experiments – CoBot robots • Using the CoBot platform • Their task catalog • Schedules work in the actual robot • 180 simulations • Scheduling times • Quality Modeling Task Monitoring and Rescheduling for Experiments 16 /20 Opportunity and Failure Management Conclusions
Experiments – Schedules • Task decomposition allows to optimize locations Schedule 1 Schedule 2 Schedule 3 Task Start End Task Start End Task Start End … 0 10 … 0 10 … 0 10 C1a 11 20 C1a 11 20 C1a 11 20 C2a 21 26 C2a 21 26 VIP 21 23 C1b 27 31 C1b 27 31 C2a 24 29 C2b 32 33 C2b 32 33 C1b 30 34 C3a 34 42 VIP 34 39 C2b 35 36 C3b 43 47 C3a 40 45 C3a 37 45 VIP 48 53 C3b 46 50 C3b 46 50 Cost 739 Cost 605 Cost 454 Modeling Task Monitoring and Rescheduling for Experiments 17 /20 Opportunity and Failure Management Conclusions
Experiments – Solving time vs. Subtasks 30 10 s, 4.4% tol. (suboptimal) Avg. solving time (s) 25 10 s 20 30 s 15 10 5 0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Task pool size (set B) (all with solutions) • Proven optimal solutions found up to size 10 Modeling Task Monitoring and Rescheduling for Experiments 18 /20 Opportunity and Failure Management Conclusions
Experiments – Quality vs. Subtasks 850 10 s, 4.4% tol. (suboptimal) 800 Avg. quality (min) 10 s 750 30 s 700 650 600 550 500 0 1 2 3 4 5 6 7 8 9 101112131415 Task pool size (set B) (all with solutions) • Quality in “10s suboptimal” is acceptable for the CoBot’s domain Modeling Task Monitoring and Rescheduling for Experiments 19 /20 Opportunity and Failure Management Conclusions
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 Experiments Task Monitoring and Rescheduling for Conclusions 20 /20 Opportunity and Failure Management
IntEx Workshop 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
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 22 /20 Opportunity and Failure Management
Recommend
More recommend
