PowerPoint Presentation
Development of Scheduling, Path Planning and Resource Management
Algorithms for Robotic Fully-automated Multi-story Parking
StructureJayanta Kumar Debnath20 July 2016University of Toledo
Electrical Engineering and Computer Science Department
Master of Science in Electrical Engineering(with concentration
in Computer Science and Engineering)
Thesis Presentation
IntroductionProblem StatementProposed MethodologyProposed Path
Planning AlgorithmProposed Elevator Scheduling AlgorithmProposed
Resource Management AlgorithmReal-time Concurrent Simulation
ModelSimulation ResultsConclusions and Future Study
Contents
IntroductionWhy Automated Parking?Create spaceIncrease
revenueBetter customer securityGreen parking50% less real-estate
than traditional parking lot!No drive way roads like traditional
parking space!Lower deployment cost plus more revenue generating
space!.More security for people and their vehicles!Increase green
space and reduce traffic congestion and carbon footprint!Promising
and effective solution for busy metropolitan areas parking
challenge!
IntroductionCurrent Automated Parking Technology
Stacker Type Automated ParkingTower Type Automated
ParkingRequire dedicated lane for stacker crane.Low space
utilization.Do not require sophisticated AI Algorithms. Stacker
Crane
Require one elevator. Low space utilization.Do not require
sophisticated AI Algorithms.Parking capacity is not scalable.
IntroductionCurrent Automated Parking TechnologyChess Type
Automated ParkingPuzzle Type Automated ParkingNo dedicated driveway
or lane required.Vehicles can be moved horizontally only. Multiple
elevators.Maximum space utilization possible.Require sophisticated
AI Algorithms.Parking capacity is highly scalable. Each cell
require lifting mechanism.Vehicles can be moved both vertically and
horizontally.No elevator required.Space utilization less than chess
type.Require sophisticated AI Algorithms.Parking capacity is highly
scalable.
Puzzle parking structure is similar to chess parking except
there is no elevators!
IntroductionCurrent Automated Parking Technology
Moving Vehicles using Roller Bed Mounted Pallets on Chess Type
Parking!
Problem StatementMotivation of this Thesis!Robotic and fully
automated parking structures are becoming increasingly feasible
from the technology perspective.There is a lack of reported designs
in literature for a computerized management system for such
structures.Artificial Intelligence is well suited for and can
enhance efficiency, scalability and mass level commercialization of
robotic fully-automated parking structures.
Problem Statement: Design, develop and prototype in simulation
an integrated software implementation for a management system that
can plan multiple concurrent paths, schedule a group of elevators,
and allocate parking space and other related resources in real time
with service times acceptable to users.
Problem SpecificationRequirements
Ground Floor Layout (shown for 1020 topology)Vehicle Movement
DirectionsNo driving lanes on any floor of the multistory parking
structureNumber of parking spaces on a given floor and number of
stories are variables.Minimum 80% utilization rate for parking on a
given floorNo more than 5 minutes waiting time for delivery or
retrieval of a vehicle by driversMultiple independent lifts (or
elevators) Robotic carts or pallets move vehicles.Unlimited number
of vehicles in motion throughout the structure at any given time
Vehicle cart and elevator movements are modeled in compliance with
physics.
Proposed MethodologyStorage ProcessStorage Management Algorithm
assigns a storage location for a Storage RequestElevator Scheduling
Algorithm assigns an elevator and informs customer.Customer leaves
their vehicle in the elevator and leave.Elevator transport vehicle
to desired floor and unload the vehicle.Path Planning Algorithm
finds a path to storage location and moves the vehicle
accordingly.
Storage and Retrieval Request Entry Kiosk!Elevator Needed?Select
Vehicle Exchange Bay and notify customerCustomer drops off their
vehicle in the Vehicle Exchange Bay and leaves.YesNo
Proposed MethodologyRetrieval ProcessStorage Management
Algorithm locates vehicle parked at a specific location for a
Retrieval Request. Elevator Scheduling Algorithm assigns an
elevator and notifies customer.Customer picks up the vehicle and
leaves.Elevator transports vehicle to ground floor.Path Planning
Algorithm finds a path towards elevator location and moves the
vehicle accordingly.
Storage and Retrieval Request Entry Kiosk!Elevator Needed?Select
Vehicle Exchange Bay and notify customerPath Planning Algorithm
finds a path towards Vehicle Exchange Bay location and moves the
vehicle accordingly.YesNo
Proposed MethodologyTheoretical Bounds
What is the minimum number of elevators?What is the minimum
number of blank cells?Bounds on minimum number of elevators and
blank cells are derivedapplying Queueing Theory on Storage and
Retrieval Processes.
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells Immovable Obstacle!Movable Obstacle!Starting CellDestination
CellDynamic EnvironmentUnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells Neighbor Cell on Planned PathUnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells Nearest blank cell located using Uniform Cost
Search!UnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells Neighbor cell on planned path of blank cellSelected blank
cellUnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells Selected blank cell moved!UnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells Vehicle moved!Check for change of immovable obstacle
topology!UnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells No change in immovable obstacle topology.Follow previously
planned path
Check for change of immovable obstacle topology.
UnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells No change in immovable obstacle topology.Follow previously
planned path
Check for change of immovable obstacle topology.
UnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells Immovable obstacle topology changed!!Re-plan path efficiently
which is a special feature of D* Lite algorithm!!
Check for change in immovable obstacle topology.
UnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells No change in immovable obstacle topology.Follow previously
planned path
Check for change in immovable obstacle topology.
UnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells No change in immovable obstacle topology.Follow previously
planned path
Check for change in immovable obstacle topology.
UnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells No change in immovable obstacle topology.Follow previously
planned path
Check for change in immovable obstacle topology.
UnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells No change in immovable obstacle topology.Follow previously
planned path
Check for change in immovable obstacle topology.
UnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells No change in immovable obstacle topology.Follow previously
planned path
Check for change in immovable obstacle topology.
UnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells No change in immovable obstacle topology.Follow previously
planned path
Check for change in immovable obstacle topology.
UnblockProcedure
Proposed Path Planning AlgorithmOverviewD* Lite AlgorithmPath
Planning for storage and retrieval of vehicles on cart.Uniform Cost
SearchLocating blank cellsD* Lite AlgorithmPath Planning for blank
cells No change in immovable obstacle topology.Follow previously
planned path
Check for change in immovable obstacle topology.
UnblockProcedure
Proposed Path Planning AlgorithmD* Lite AlgorithmD* Lite
Algorithm is Heuristic Based and Incremental fast re-planning
search algorithm and very effective in a dynamic environment.
Uninformed breadth-first searchHeuristic based D* Lite or A*
searchHeuristic is good!Grey cells are explored during path
planning.
Proposed Path Planning AlgorithmD* Lite AlgorithmD* Lite
Algorithm is a Heuristic Based and Incremental fast re-planning
search algorithm and very effective in a dynamic environment. A*
searchIncremental D* Lite searchIncremental search is able to
effectively re-use partial path plan from previous search!
Obstacle added!Obstacle added!Grey cells are explored in
re-planning.
Proposed Path Planning AlgorithmUnblock Procedure
Use uniform cost search to locate nearest blank cell
Proposed Path Planning AlgorithmUnblock ProcedureUniform cost
search 1st Iteration
No blank cell Found!
Proposed Path Planning AlgorithmUnblock ProcedureUniform cost
search 2nd IterationNo blank cell Found!
Proposed Path Planning AlgorithmUnblock ProcedureUniform cost
search 3rd IterationNo blank cell found!
Proposed Path Planning AlgorithmUnblock ProcedureUniform cost
search 4th IterationBlank cell found!
Proposed Path Planning AlgorithmUnblock Procedure
Two blank cells! Two possible destination cells!Find nearest
blank cell destination cell pair.
Proposed Path Planning AlgorithmUnblock Procedure
Nearest blank cell destination cell pair Find nearest blank cell
destination cell pair.
Proposed Path Planning AlgorithmUnblock ProcedureUse D* Lite
path planning to move blank cell towards destination cell
Proposed Elevator Scheduling AlgorithmTwo-level Integer
Programming Formulation Problem Formulation
Each trip will serve one vehicle!HNPGA (Hybrid Nested Partition
and Genetic Algorithm) used for High Level Assignment!FIFO used for
Low Level Assignment!
Proposed Elevator Scheduling AlgorithmHNPGA : Select Next Most
Promising RegionTwo steps at each iteration of HNPGA for selecting
next Most Promising Region Step 1: Select best sub region.Step 2:
Global verification of selected best sub region with Surrounding
Regions.Both steps use Genetic AlgorithmNested Partition Tree!
Proposed Elevator Scheduling AlgorithmStep 1: Select Best Sub
RegionTotal 10 vehicles; Each field represents vehicles to schedule
Initial populations of GAAfter evaluation cycles!Fittest of final
populationsCrossover and MutationBest sub region found by GA!
Selected Most Promising Region at first iteration.
40
Proposed Elevator Scheduling AlgorithmStep 2: Global
VerificationInitial populations of GA uniformly taken from three
region.Selected best sub region! Selected Most Promising Region at
first iteration.
41
Proposed Elevator Scheduling AlgorithmObjective Function of
GA
Each field or gene of chromosome represents vehicles to
schedule. The value of each field represents the assigned
elevator.Total time required to complete storage or retrieval
process associated with assigned vehicles.
Proposed Elevator Scheduling AlgorithmCrossover Operator
Proposed Elevator Scheduling AlgorithmMutation Operator
Theoretical Bounds on Resource NeedsQueueing Theory
Server - 01Server - 02Server - S
Mean Arrival Rate, Steady State Condition!M/M/S Queue Model
Stochastic Arrival ProcessStochastic Service ProcessMultiple
Parallel ServersWithout steady state queue will grow infinitely
large eventually.
Proposed Resource Management AlgorithmStatistical ModelsRush
hour period customer arrival modelingMorning Rush Hour2 clock hour
period from 6:30 AM to 8:30 AM95% of requests are storage5% of
requests are retrievalEvening Rush Hour2 clock hour period from
4:00 PM to 6:00 PM95% of requests are retrieval5% of requests are
storageInspired by busy downtown business districts traffic
pattern
Proposed Resource Management AlgorithmStatistical ModelsPoisson
Distributed Customer Arrivals with varying mean arrival rate!
Theoretical Bounds on Resource NeedsBound on Minimum Number of
Blank Cells
Modeled part of retrieval process at the beginning of evening
rush hour as M/M/S queue where blank cells act as multiple servers
to transport vehicles toward elevator load/unload bay!
Theoretical Bounds on Resource NeedsBound on Minimum Number of
Blank Cells
Theoretical Bounds on Resource NeedsBound on Minimum Number of
Elevators
Modeled part of storage/retrieval process as M/M/S queue where
elevators act as multiple servers to transport vehicles between
floors!
Theoretical Bounds on Resource NeedsBound on Minimum Number of
Elevators
Starting floorDestination floor DistanceSlow downSpeed
upElevator Dynamics
Theoretical Bounds on Resource NeedsBound on Minimum Number of
Elevators
Real-time Concurrent Simulation ModelUnified Modeling Language
The overall functionality of simulation is modeled through five
major activity modules, which are (a) Automated Parking Lot, (b)
Automated Storage Controller, (c) Automated Retrieval Controller,
(d) Elevator Controller, and (e) Elevator Scheduler
Modular Simulation Architecture
Real-time Concurrent Simulation ModelUnified Modeling Language
State machine diagram for Automated Retrieval Controller: moving
towards elevators
Real-time Concurrent Simulation ModelUnified Modeling Language
State machine diagram for Automated Storage Controller: moving from
elevators
Real-time Concurrent Simulation ModelUnified Modeling Language
State machine diagram for Elevator Controller : moving between
floors
Real-time Concurrent Simulation ModelUnified Modeling Language
Busy-wait Synchronization Techniques used to communicate among
concurrent threads.
Timing diagram
Simulation StudyExperimental SetupSpace Utilization > 80% The
capacity of parking lot needs to be fully utilized within two
clock-hour period43 Test Cases Found!Generating Test CasesNumber of
columns on each floor layoutNumber of rows on each floor
layoutNumber of floorsMaximum value of mean arrival rate for
vehicle requests for the entire parking structure per hour among
all rush hour time slots
Simulation StudySimulation SoftwareA software application with
multithreading was developed through the Unified Modeling Language
(UML) using Java and MATLAB programming languages. Simulation
Software was run in Linux environment for better multithreading
capability!
Simulation StudyRunning Simulation in 50X Speedup
Simulation StudySimulation ResultsAverage customer waiting time
within an impressive 5 minutes mark!
Simulation StudySimulation ResultsIn most cases, average
customer waiting time within an impressive 2 minutes mark!
Simulation StudySimulation ResultsFor most cases, the maximum
(worst case) customer waiting time is less than 5 minutes although
for a small number of cases it was between 10 to 17 minutes!
Simulation StudySimulation ResultsExtreme maximum values occur
for very few test cases with 10% immovable carts. In general,
maximum waiting times are within the 6-minute mark. Low
Probability!
Simulation StudySimulation Results for Case #42 with 10%
Immovable CartsMost of the customers experience average waiting
times; very few customers have to wait more than the average
value!Frequency Distribution of Waiting Times for Individual
customers for Case #42Considering 10% immovable cartsConsidering
10% immovable carts
ConclusionsConclusionsIn light of and within the context of the
simulation study presented, the design appears feasible for real
time deployment in an industrial-grade environment.Average Customer
Waiting Time is not more than 5 minutes in most cases! Space
Utilization for parking is more than 80% !Design supports customer
arrival rates of up to 800 customers per hour!
Future StudyRecommendationsIn the current system, all the
parking spots are the same size. Given that there are different
size vehicles (sedans, SUVs, mini vans, trucks, etc.) to park, the
size of a parking spot would have to match the largest car size. To
maximize the available real-estate space utilization rate and
enhance the capacity, future studies may consider the design other
topologies which may have different size parking spaces.
Study of the effect of robotic cart failures could be extended
further to determine the adverse impact on performance more
closely.
We assumed, for the analysis based on the queueing theory, that
customers would not engage in balking or reneging in the waiting
lines. In future studies these and other similar complications can
be injected into the statistical models to determine their effects
on performance.The research could be extended in the future from
the following aspects:
PublicationsDebnath, Jayanta K., and Gursel Serpen. "Real-Time
Optimal Scheduling of a Group of Elevators in a Multi-Story Robotic
Fully-Automated Parking Structure." Procedia Computer Science 61
(2015): 507-514.
J. Debnath and G. Serpen, Design of Multithreaded Simulation
Software through UML for a Fully Automated Robotic Parking
Structure, to appear in proceedings of International Conference on
Simulation Modeling Practice and Theory, Las Vegas, Nevada, July
2016.
Thank you!Any Questions?