Dynamic Temporal Planning for Multirobot Systems

Dynamic Temporal Planning for

Multirobot Systems

C. Ugur Usug and Sanem Sarıel-Talay

{usugc,sariel}@itu.edu.tr

Outline

•

Motivation

–

Multirobot cooperation (earlier work)

•

Failure Recovery

•

Dynamic Temporal Planning Framework

–

Proposed methods

•

Preliminary Results

DEMiR-CF:

Distributed and Efficient Multi-Robot-Coop Framework

•

D

istributed &

E

fficient

M

ult

i

-

R

obot -

C

ooperation

F

ramework

•

A distributed multi-robot cooperation framework

–

for autonomous robots to achieve a complex mission in cooperation

–

Each robot selects its most suitable task in a

distributed manner

•

Rough schedules

are formed to consider the problem as a whole

•

Dynamic task selection

is carried out according to rough schedules

–

Appropriate robots are

allocated

to the selected tasks

in a

distributed

manner

•

Contingencies

are

handled

in real time

–

Recovery routines are applied for correcting models

(*) A Generic Framework for Distributed Multirobot Cooperation, Sanem Sariel-Talay, Tucker Balch and Nadia

Erdogan ,

Journal of Intelligent & Robotic Systems, 2011, Vol 63(2), pp. 323-358

Distributed Coordination

R1: move, drop

R2: move, drop

R3: move, drop, clean

R4: move, clean

The focus of this study: Failure Recovery

•

Execution Failures due to uncertain information

–

Replanning does not resolve the problem

–

Additional actions are needed

Valid plan – High level

Execution Failures:: Permanent Failure

•

Permanent Failure

–

resources are not adequate to move the object

–

cannot be resolved

R1

Wall

B1

Target

(move-to-loc ?robot ?location)

(move-to-obj ?robot ? location)

(pick ?robot ?smallObject)

(drop ?robot ?smallObject)

(push ?robot ?largeObject)

Execution Failures:: Temporary Failure

•

Temporary Failure

–

Unstructured environment: The robot is not aware of

the obstacle

–

Replan?

Execution Failures:: Temporary Failure

•

Temporary Failure

–

Unstructured environment: The robot is not aware of

the obstacle

–

Replan?

Execution Failures:: Temporary Failure

•

Temporary Failure

–

Unstructured environment: The robot is not aware of

the obstacle

–

The operator “push” is defined in the domain

R1 B2

Wall Wall

B1

Recovery of Failures at the high level

•

Deactivate/Suspend failed actions

–

Till when?

•

Update the problem and the domain (inc. start

and end states)

–

Reasoning?

–

If definitions are not adequate (abstraction)?

•

Continual planning and subgoal planning

–

Subgoal determination?

Laziness in Representation

(def-adl-operator (push ?robot ?obj ?t ?s)

(pre

(?robot) (agent ?robot)

(?obj) (large-object ?obj)

(?t) (push-time ?t)

(?s) (push-speed ?s)

(?c) (hcost ?c)

(and

(notbusy ?robot)

(empty ?robot)

(= (xcoord ?robot) (xcoord ?obj))

(= (ycoord ?robot) (ycoord ?obj))

)

Updating the Domain

•

Temporary Failure in Execution

–

Due to uncertain information

•

Assumptions

–

The cause of failure is known

–

The knowledge may not be sufficient for reasoning

–

No experts

Dynamic Temporal Planning System

The Proposed Approach

•

Suspend the failed action

–

until the failure is believed to be resolved

•

Resume the failed action

–

When the situation of the cause is changed

•

Search for domain operators that affect the cause

in case they resolve the problem

M1: Adding a pseudo operator

(def-adl-operator (move-to-loc ?loc)

(pre

…

(not (move-to-loc

_

locked ?loc))

)

…

)

(def-adl-operator (

pseudo1

?obj ?loc)

(pre

(and

(= ?obj obstacle)

(move-to-loc

_

locked ?loc)

(or

(not (xcoord ?obj obj

_

x))

(not (ycoord ?obj obj_y))

)

)

)

(del

(move-to-loc

_

locked ?loc))

Properties of the

cause of failure

Failed Action

Related to the

failed action

Related to the

cause of the failure

M2: Adding replicated operators

(def-adl-operator (move-to-loc ?loc)

(pre

…

(not (move-to-loc

_

locked ?loc))

)

…

)

(def-adl-operator (

push-2

?obj ?loc)

(pre

…

(and

(= ?obj obstacle)

(move-to-loc

_

locked ?loc)

(or

(not (xcoord ?obj obj

_

x))

(not (ycoord ?obj obj_y))

)

M3: Adding a replicated failed operator

(def-adl-operator (move-to-loc ?loc)

(pre

…

(not (move-to-loc

_

locked ?loc))

)

…

)

(def-adl-operator (

move-to-loc-2

?loc ?obj)

(pre

…

(and

(= ?obj obstacle)

(move-to-loc

_

locked ?loc)

(or

(not (xcoord ?obj obj

_

x))

(not (ycoord ?obj obj_y))

)

)

)

…

)

(move-to-loc ?robot ?location)

(move-to-obj ?robot ? location)

(pick ?robot ?smallObject)

Experiment I – Settings

Experiment I – Results

–

2,53 GHz / 4GB RAM

–

Max Time: 500s

–

Memory: 20000 nodes

#GN #EN Plan Time (s) Makespan(s) #GN #EN Plan Time (s) Makespan(s)

1 robot 1 box M1 46 23 0,038 46,97 103 54 0,088 46,97 M2 N/A 123 68 0,116 46,97 M3 44 22 0,04 46,97 79 45 0,078 46,97 1 robot 2 box M1 105 45 0,11 96,77 1148 550 0,78 81,01 M2 N/A 1050 558 0,8 81,01 M3 100 44 0,106 96,77 782 421 0,868 81,01 1 robot 3 box M1 177 70 0,194 143,38 21586 9720 59,25 118,4 M2 N/A 19569 9779 50,55 118,4 M3 175 69 0,194 143,38 14794 7425 25,3 118,4 2 robot 2 box M1 128 48 0,136 76,08 35193 12202 170,1 45,13 M2 N/A 38128 14376 195,13 45,13 M3 133 50 0,144 76,08 26793 10328 75,58 45,13 2 robot 3 box M1 319 101 0,323 151,19 N/A M2 N/A N/A M3 296 97 0,31 151,19 N/A 2 robot 4 box M1 609 172 0,586 240,43 N/A M2 N/A N/A M3 618 174 0,634 240,43 N/A

Simulations – Settings

TLPlan Output File Domain File

TLPlan Manager

TLPlan Manager

Problem & Domain Parser

Problem & Domain Parser

Simulations – Videos

initial

plan

new

plan

Conclusion

•

Dynamic Temporal Planning System

–

Temporary failures can be handled

•

Domain updates

–

Ensure the planner select alternative paths

•

Future work

–

Real robot experiments

–

Reasoning about the failures

Questions?

C. Ugur Usug and Sanem Sarıel-Talay

Istanbul Technical University

Artificial Intellingence and Robotics Laboratory

http://air.cs.itu.edu.tr