Action Preparation and Replanning in Manipulation

Hisashi Hayashi

and Hideki Ogawa

Corporate Research & Development Center, Toshiba Corporation, 1 Komukai-toshiba-cho, Saiwai-ku, Kawasaki, Japan

Corporate Manufacturing Engineering Center, Toshiba Corporation, 33 Sin-isogo-cho, Isogo-ku, Yokohama, Japan

Keywords:

Speculative Action Preparation, Replanning, Manipulation, Robotics.

Abstract:

In order to pick (place) a target object from (on) a shelf, a service robot moves to the front side of the shelf,

removes obstacles, and reaches out a hand. If the robot prepares for the next arm manipulation while moving

to the shelf, it is possible to save time for plan execution. The robot also needs to replan if, for example, a

person removes obstacles for the robot. After replanning, the robot might need to suspend the current action

execution or next action preparation before executing the updated plan. This paper introduces a method to

integrate planning, action execution, speculative next action preparation, replanning, and action suspension

based on Hierarchical Task Network (HTN) planning. We also show that this method is effective for pick-and-

place manipulation in dynamic environments.

1 INTRODUCTION

Pick-and-place manipulation in dynamic environ-

ments is an important research theme in robotics.

Typically, a service robot moves around the environ-

ment, picks an object from a shelf, and places the ob-

ject on the target shelf. The robot needs to use a right

or left hand to pick or place the object without colli-

sion with obstacles or removes the obstacles before-

hand.

In order to implement service robots of this

type, combinations (Cambon et al., 2009; Choi and

Amir, 2009; Haspalamutgil et al., 2010; Hauser and

Latombe, 2009; Kaelbling and Lozano-Perez, 2010;

Wolfe et al., 2010) of a high-level general-purpose

task planner and a low-level motion planner such as

RRT (LaValle and Kuffner, 2001) have played impor-

tant roles. Given a manipulation task, the high-level

task planner decides which object to move and where

to bring the object. The low-level motion planner con-

trols the arm, considering kinematic constraints and

geometric map information. Manipulation planning

“among movable obstacles” is also implemented by

combinations (Nieuwenhuisen et al., 2008; Ota, 2009;

Stilman et al., 2007; Zacharias et al., 2006) of a high-

level algorithm and a low-level motion planner.

A typical task plan is “move to table1”, “pick ob-

ject1”, “move to table2”, and “place object1”. The

execution times of these actions are long. In order to

shorten the execution time of the plan, it is effective

to move an arm while moving to a shelf and prepare

for the next pick (or place) action.

On the other hand, a service robot working in the

real world needs to replan when the environment un-

expectedly changes. For example, before the robot

removes an obstacle, a person might remove it for the

robot before the robot arrives in front of the shelf. In

this case, the robot no longer needs to remove the ob-

stacle. Therefore, the robot needs to replan and cancel

the action preparation for removing the obstacle.

In order to adapt to a dynamic environment, we

understand that replanning is necessary. In addition,

the following new functions are necessary:

• Integration of action preparation and replanning;

• Suspension of preparing/executing actions.

In this paper, we extend an online Hierarchi-

cal Task Network (HTN) planning agent Dyna-

gent

(Hayashi et al., 2006) so that it integrates ac-

tion preparation and suspension of action prepara-

tion/execution. We also show that these new functions

of online planning are effective for service robots

that pick and place objects in dynamic environments.

Note that in the previous work of Dynagent the ef-

fects of replanning and suspension of action execu-

tion/preparation in the pick-and-place manipulation

have not been investigated.

Dynagent is a registered trademark of Toshiba Corpo-

ration.

109

Hayashi H. and Ogawa H..

Action Preparation and Replanning in Manipulation.

DOI: 10.5220/0005152101090116

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence (ICAART-2015), pages 109-116

ISBN: 978-989-758-073-4

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

The rest of this paper is organized as follows: Sec-

tion 2 discusses related work. Section 3 brieﬂy ex-

plains the online HTN planning algorithm that we use

in this work. Section 4 brieﬂy explains our heuris-

tics for object manipulation. Section 5 shows how

to prepare for future action executions using existing

online HTN planners. Section 6 shows how to report

the result of an action execution and update the plans.

Section 7 shows how to suspend current action execu-

tion/preparation after replanning. Section 8 presents

the result of an experiment. Section 9 presents the

result of another experiment. Section 10 is the con-

clusion.

2 RELATED WORK

Some research on replanning in robotics has been re-

ported. In (Philip et al., 2009), a motion planner

makes plans to control a humanoid robot. Using vi-

sion information, this planner dynamically updates

the path. In another paper (Hayashi et al., 2009), an

HTN planner is used to compute the path of a mobile

robot. While executing a plan, the planner controls

the sensing target in the background to obtain new in-

formation that might hinder the plan execution. How-

ever, in none of this research, do the robots prepare

for the next action execution while executing another

action.

Motion planning among moving obstacles or hu-

mans is an important research topic. In (Mainprice

and Berenson, 2014), prediction of human motions,

replanning, and execution are interleaved. At the re-

planning step, the next motion plan is made while the

robot is moving, which is important for saving time.

Because humans move, their locations are predicted

before planning to avoid collisions. In another ap-

proach (Park et al., 2013), GPU is used to compute

the trajectory of the robot in parallel. They also inter-

leave planning and execution. Because the planning

times become short, the robot can quickly respond to

the dynamic environment. In none of this research,

however, is the next action executed in parallel with

the current action. In addition, unlike in the present

work, unnecessary action execution is not suspended

after replanning. Furthermore, they do not use real

robots for evaluation.

Replanning has a long history of research in ar-

tiﬁcial intelligence. A well-known planner, STRIPS

(Fikes and Nilsson, 1971), uses a triangle table (Fikes

et al., 1972) to check the preconditions of actions in

the plan. As explained in (Russell and Norvig, 1995),

the ﬁrst partial-order online planner that smoothly in-

tegrates planning, replanning, and execution is IPEM

(Ambros-Ingerson and Steel, 1988), which uses “pro-

tected links” for precondition checking. Protected

links were ﬁrst introduced in NONLIN (Tate, 1977)

and used in many partial-order planners, including

TWEAK (Chapman, 1987) and SNLP (McAllester

and Rosenblitt, 1991). However, unlike in the present

work, these techniques are not used to suspend unnec-

essary action execution after replanning.

Parallel plan execution is closely related to next

action preparation since multiple actions are executed

in both approaches. In (Jacobs et al., 2012), motion

plans are executed in parallel. They do not use a high-

level planner such as an HTN planner. In (Einig et al.,

2013), an architecture of parallel plan execution for

mobile service robots is presented. They use a partial-

order HTN planner called JSHOP2 (Nau et al., 2003).

They also claim that they can extend their system so

that the robot can replan when action execution fails.

Our motivation is the same in that both parallel action

execution and replanning would reduce the total plan

execution time in manipulation. However, they have

not yet evaluated the effect of replanning.

3 ONLINE HTN PLANNING

We use Dynagent (Hayashi et al., 2006), a general-

purpose forward-chaining HTN planning agent work-

ing in a dynamic environment. Although we use it for

pick-and-place manipulation, it can also be used for

other purposes. Forward-chainingHTN planners such

as Dynagent and SHOP (Nau et al., 1999) make plans

by incrementally decomposing an abstract task into

more detailed tasks in the same order that they will be

executed, and search for an optimal/suboptimal plan

by best-ﬁrst search. Rule programmers of HTN plan-

ners can express domain control heuristics using task

decomposition rules.

The agent does not have to decompose an abstract

task till the agent needs to execute the task. (This is

called “lazy evaluation” or “lazy planning”.) How-

ever, the ﬁrst task in each plan must be an executable

action before selecting the plan for execution.

It keeps and modiﬁes several alternative plans

while executing a plan. Therefore, it can change the

plan to an alternative plan when the current plan be-

comes invalid or another plan becomes more attrac-

tive in terms of costs.

After successfully executing an action, the agent

updates the belief based on the effect of the action,

and removes the executed action from each plan if the

ﬁrst action in the plan is uniﬁable

with the executed

Two actions are uniﬁable if and only if they become

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

110

action. When the agent fails to execute an action, it

removes each plan such that the ﬁrst action in the plan

is uniﬁable with the action that could not be executed.

Invalid plans can be found by rechecking the (pro-

tected) precondition of each action in the plans. Some

preconditions must be protected if they might become

unsatisﬁable when the belief is updated. Even if the

precondition of an action is unsatisﬁable when plan-

ning initially, it might become satisﬁable later be-

cause of a belief update. Therefore, the agent keeps

such invalid plans as well.

4 MANIPULATION STRATEGY

This section brieﬂy explains our heuristics for object

manipulation that was shown in our previous paper

(Hayashi et al., 2013) in more detail. We express the

heuristics using the task decomposition rules of HTN

planning. The top-level task is to bring object1 from

shelf1 to shelf2. In order to execute this task, we ex-

ecute the following plan (a sequence of subtasks): 1.

move to shelf1; 2. pick object1 with the right (left)

hand; 3. move to shelf2; 4. place object1 with the

right (left) hand.

In order to pick or place the object with the right

(left) hand, the robot stretches the arm diagonally

from the right (left) front of the shelf. Therefore, all

the obstacles on the right (left) front of the shelf have

to be removed before stretching the arm to avoid colli-

sions. In order to place an object on a shelf, the robot

places the object as far as possible from the robot so

that the robot can place many objects on the shelf.

There are many ways to remove or avoid obsta-

cles. Best-ﬁrst search ﬁnds a low-cost plan. In other

words, it tries to minimize the number of obstacle re-

movals.

5 ACTION PREPARATION

In order to prepare for the next manipulation while

moving to a shelf, it is necessary to know the next

actions to execute after the current action execution.

For this purpose, we add an argument for each action

so that it can record the list of next actions.

For example, suppose that the agent makes the fol-

lowing plan:

1. goto(shelf1);

2. pickL(object1,shelf1).

syntactically the same by substituting some variables.

In this plan, the robot goes to shelf1 and picks object1

with the left hand.

The agent ﬁrst tells the lower-level module to ex-

ecute goto(shelf1). However, the lower-level module

does not know the next action after goto(shelf1). For

this reason, we modify the task-decomposition rules

of Dynagent so that the following plan is made:

1. goto(shelf1,[pickL(object1,shelf1)]);

2. pickL(object1,shelf1.[]).

In this case, the agent tells the lower-level module

to execute goto(shelf1,[pickL(object1,shelf1)]). Now,

the lower-level module can prepare for the execution

of pickL(object1,shelf1) while executing goto(shelf1).

In this way, we do not have to change the framework

or the source code of Dynagent.

In general, this strategy is summarized as follows:

Strategy 1. (Planning for action preparation)

Consider the following plan:

1. a

, ..., t

)

2. a

, ..., t

)

3. a

n−1

, ..., t

n−1

)

4. a

, ..., t

)

In order to give the information of the next two

actions to the lower-levelaction execution module, we

modify the task-decomposition rules of Dynagent so

that the following plan is made:

1. a

, ..., t

, [a

, ..., t

)])

2. a

, ..., t

, [a

, ..., t

)])

3. a

n−1

, ..., t

n−1

, [a

, ..., t

)])

4. a

, ..., t

, [])

In this strategy, each action records only the next

action. In the same way, it is possible to record mul-

tiple actions if necessary.

6 REPORTING RESULTS

Strategy 1 in the previous section works as long

as the plan is not modiﬁed in the middle of the

action execution. However, if we need to re-

plan, we have another problem. In the example

of the previous section, after the robot moves to

shelf1, the low-level action executor reports to Dy-

nagent that goto(shelf1,[pickL(object1,shelf1)]) has

been successfully executed. Then, Dynagent re-

moves goto(shelf1,[pickL(object1,shelf1)]) from the

plan. However, Dynagent cannot remove the ﬁrst ac-

tion from the following alternative plan:

ActionPreparationandReplanninginManipulation

111

1. goto(shelf1, [pickR(object2,shelf1)])

2. pickR(object2,shelf1,

[placeR(object2,shelf1,cell7)])

3. placeR(object2,shelf1,cell7,

[pickL(object1,shelf1)])

4. pickL(object1,shelf1, [])

In this alternative plan, the robot goes to shelf1, picks

object2 with the right hand, places object2 at cell7 on

shelf1 with the right hand, and picks object1 with the

left hand.

Here, the ﬁrst action is the same as

goto(shelf1,[pickL(object1,shelf1)]) except that

the next action is different. Because Dynagent is a

general-purpose planning agent, it recognizes that

goto(shelf1,[pickL(object1,shelf1)]) and goto(shelf1,

[pickR(object2,shelf1)]) are different actions because

they are not uniﬁable.

This problem can be avoided if the low-level ac-

tion executor reports to Dynagent that it has executed

goto(shelf1,

) where is a new variable. In this

way, both goto(shelf1,[pickL(object1,shelf1)]) and

goto(shelf1, [pickR(object2,shelf1)]) become uniﬁ-

able with goto(shelf1,

). Therefore, these ﬁrst actions

are removed from both of the plans.

Similarly, when failing to execute this action, if

the low-level action executor reports to Dynagent that

it has failed to execute goto(shelf1,

), then Dynagent

removes these two invalid plans.

In general, this strategy is summarized as follows:

Strategy 2. (Reporting the action execution result)

After executing the ﬁrst action:

, ..., t

, [a

, ..., t

)])

of the plan that was made by Strategy 1, it reports the

execution result (success or failure) of the following

form of the action:

, ..., t

)

where

is a new variable

In this way, we do not have to change the frame-

work or the source code of Dynagent.

7 ACTION SUSPENSION

Dynagent changes the plan as the situation changes.

Sometimes it becomes unnecessary to ﬁnish the cur-

rent action execution if the current plan becomes in-

valid. In robotics, the execution time of one action

is long and it is better to suspend such unnecessary

action execution to save time. Sometimes the robot

might ﬁnd a new obstacle while picking or placing an

object. In this case, it is necessary to stop stretching

the arm to avoid collision with the obstacle.

Similarly, the current action preparation might be-

come unnecessary after modifying the plan. In this

case, we would like to stop the current action prepa-

ration as well.

From these view points, we suspend unnecessary

action execution and action preparation as follows:

Strategy 3. (Replanningand suspension of action ex-

ecution/preparation)

When the belief is updated, plans are modiﬁed and

the action execution/preparation is suspended as fol-

lows:

1. Belief

update instruction is given as an input.

2. Update the belief as instructed.

3. If a plan is being executed, do the following pro-

cedure:

(a) If the current action execution/preparationcan-

not be suspended, wait until the current action

execution/preparation ﬁnishes.

(b) Modify the plans

including alternative plans.

plans.

(d) If an action is being executed or prepared and

it becomes unnecessary, do the following pro-

cedure:

i. Send an instruction to suspend the current ac-

tion execution/preparation to the lower-level

action execution module.

ii. Update the belief considering the effect

of the suspension of the action execu-

tion/preparation.

iii. Remodify the plans including alternative

plans.

iv. Rechoose a plan for execution from alterna-

tive plans.

(e) Resume the plan execution.

8 EXPERIMENT 1

As shown in Figure 1, in the initial state, there exist

the objects A, B, and C respectively at cells 1, 5, and

6 on shelf1. Each object has an RFID tag and rough

location on the shelf is recognized through the RFID

reader attached underneath.

Belief is the knowledge that is used for planning.

Although we use the replanning algorithm of Dyna-

gent, any replanning algorithm can be used in this strategy.

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

112

7 8 9

shelf1

(Side Table)

2. Right Hand

3. Left Hand

Robot

1. Approach

Figure 1: Initial Plan.

We used a service robot called SmartPal V

(Yaskawa Electric Corporation, 2007). SmartPal V

is a mobile robot equipped with two hands/arms. The

robot can recognize an object through a stereo camera

and pick the recognized object.

We installed a forward-chaining HTN planning

agent, Dynagent on a PC (Windows XP), and con-

nected it with other modules on SmartPal V through

the middleware called OpenRTM-aist 1.0.0.

In the following subsections, we ﬁrst explain the

three scenarios that were used for experiments and

then show the results of the experiments.

8.1 Static Scenario 1

Given the task of picking B, the robot ﬁrst approaches

shelf1. While approaching shelf1, the robot raises

the right arm to prepare for the next action execution.

When the robot stops in front of cell 6, it recognizes

C with the stereo camera, picks C, and places it at

cell 7 with the right hand avoiding collision with B.

The robot then recognizes B and picks B with the left

hand. When the environment does not change, we

conﬁrmed that the robot works as expected as shown

in Figure 1.

8.2 Dynamic Scenario 1

This scenario is the same as static scenario 1 in

the previous subsection except that the environment

changes in the middle of the plan execution. There-

fore, this scenario includes replanning in the middle

of action execution.

In this scenario, while the robot is stretching the

right arm to try to pick obstacle C, a person removes

the obstacle on behalf of the robot as shown in Fig-

ure 2. After the robot ﬁnds that the obstacle has been

removed through the RFID reader attached under the

table, it replans, stops picking C, and directly picks

the target object B with the left hand.

SmartPal is a registered trademark of Yaskawa Electric

Corporation.

Similar static scenario is reported in more detail in our

previous paper (Hayashi et al., 2013).

Figure 2: Replanning and Action Suspension.

8.3 Dynamic Scenario 2

This scenario is similar to dynamic scenario 1 in the

previous subsection. The difference is the timing of

the help by a person. In addition, this scenario in-

cludes replanning in the middle of action execution

and the speculative next action preparation. Note that

dynamic scenario 1 in the previous subsection does

not include replanning in the middle of the “specula-

tive next action preparation”.

Figure 3: Replanning and Action Preparation Suspension.

In this scenario, a person removes the obstacle

while the robot is approaching shelf1 as shown in Fig-

ure 3(a) Note that the robot is raising the right arm

while moving to prepare for the next action. After

the robot ﬁnds that the obstacle has been removed,

it replans, stops preparing for picking C, and starts

preparing for picking the target object B while mov-

ing to shelf1. Therefore, in Figure 3(b), the robot is

putting the right arm down and raising the left arm

while moving to shelf1.

8.4 Result

This subsection shows the experimental results of the

previous subsections. We conducted experiments in

the following six patterns:

ActionPreparationandReplanninginManipulation

113

• Pattern 1a: Static scenario 1

• Pattern 2a: Dynamic scenario 1

• Pattern 3a: Dynamic scenario 2

• Pattern 1b: Same as Pattern 1a except that the

robot does not prepare for the next picking action

while approaching shelf1.

• Pattern 2b: Same as Pattern 2a except that the

robot does not prepare for the next picking action

while approaching shelf1.

• Pattern 3b: Same as Pattern 3a except that the

robot does not prepare for the next picking action

while approaching shelf1.

In Patterns 1a, 2a, and 3a, the robot prepares for the

next picking action while approaching shelf1. On the

other hand, it does not in Patterns 1b, 2b, and 3b.

We measured the execution times of the plans for

the robot to approach shelf1 and pick up the target

object B. However, the time for vision-recognition is

not included. Table 1 shows the results of the experi-

ments.

By comparing the results for patterns 1a and 1b (or

the patterns 2a and 2b, or the patterns 3a and 3b), we

understand that the robot saves 5 seconds by prepar-

ing for the next pick action while approaching shelf1.

By comparing the results for patterns 1a and 2a (or

the patterns 1b and 2b), we understand that the robot

saves 40 seconds by suspending the pick action of C

and omitting the place action of C.

By comparing the results for patterns 1a and 3a (or

the patterns 1b and 3b), we understand that the robot

saves 53 seconds by omitting the pick action and the

place action of C.

From these results, we can conﬁrm that action

preparation, replanning, and suspension of unnec-

essary action execution/preparation are effective not

only for adapting to dynamic environments but also

for saving plan execution time.

Table 1: Time for Moving and Picking the Target Object.

324585274080Time (sec)

3b2b1b3a2a1aPattern

324585274080Time (sec)

3b2b1b3a2a1aPattern

9 EXPERIMENT 2

In this experiment, we use the same robot and envi-

ronment as in Experiment 1. The robot moves around

the room shown in Figure 4. The side table in Figure

1 that was used in Experiment 1 is also used in this

experiment as shown in Figure 4. In the initial state,

the robot is at node 1. The object A exists at cell 5 on

the side table. No other objects exist.

bookcase

rack

fridge with two shelves

side table

bed table

bed

1 2 3

4 5 6

7 8 9

wall

trash can

Figure 4: Static Scenario 2 and Dynamic Scenario 3.

In the following subsections, we ﬁrst explain the

four scenarios that were used for experiments and

then show the results of the experiments.

9.1 Static Scenario 2

As shown in Figure 4, given the task to bring A from

the side table to the bookcase, the robot makes a plan,

moves from node 1 to node 2, picks A with the right

hand, moves from node 2 to node 5, places A at cell

1 in the bookcase, and ﬁnishes the task. Soon after-

wards, the robot is given another task to move A from

the bookcase to the trash can. The robot makes a plan,

picks A with the right hand, moves from node 5 to

node 6, puts A into the trash can, and ﬁnishes the sec-

ond task. In this scenario, the robot plans and com-

pletes two tasks consecutively. However, the robot

does not replan.

9.2 Dynamic Scenario 3

This scenario is similar to static scenario 2. However,

the timing of receiving the second task is different.

While placing A at cell 1 in the bookcase to com-

plete the ﬁrst task, the robot is requested to put A into

the trash can. In other words, the destination of A is

changed. Now the robot does not need to place A in

the bookcase. The robot replans, suspends the action

execution, moves from node 5 to node 6, puts A into

the trash can, and ﬁnishes the second task.

9.3 Dynamic Scenario 4

This scenario is similar to static scenario 2 and dy-

namic scenario 3. However, the timing of receiving

the second task is different. While the robot is mov-

ing from node 2 to node 5 to approach the bookcase,

as shown in Figure 5, the robot is requested to put A

into the trash can. Now the robot does not need to go

to the bookcase. The robot replans, suspends the ac-

tion execution, moves to node 6, puts A into the trash

can, and ﬁnishes the second task.

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

114

rack

fridge with two shelves

side table

bed table

bed

1 2 3

4 5 6

7 8 9

wall

trash can

bookcase

Figure 5: Dynamic Scenario 4.

9.4 Dynamic Scenario 5

This scenario is similar to static scenario 2, dynamic

scenario 3, and dynamic scenario 4. However, the

timing of receiving the second task is different. While

the robot is moving from node 1 to node 2 to approach

the side table, as shown in Figure 6, the robot is re-

quested to put A into the trash can. Now the robot

does not need to start the action to go to the bookcase.

The robot replans, continues to go to node 2, picks

A with the right hand, moves from node 5 to node 6,

puts A into the trash can, and ﬁnishes the second task.

rack

fridge with two shelves

side table

bed table

bed

1 2 3

4 5 6

7 8 9

wall

trash can

bookcase

Figure 6: Dynamic Scenario 5.

9.5 Result

This subsection shows the experimental results of the

previous subsections. We conducted experiments in

the following four patterns:

• Pattern 4: Static scenario 2

• Pattern 5: Dynamic scenario 3

• Pattern 6: Dynamic scenario 4

• Pattern 7: Dynamic scenario 5

The purpose of these experiments is to evaluate how

much time can be saved by replanning and action/plan

suspension. We expect that the robot can save time by

suspending unnecessary movements between nodes.

We measured the execution times for the robot in the

case of starting from node 1 and putting A into the

trash can. However, the time for vision-recognition is

not included. Table 2 shows the results of the experi-

ments.

By comparing the results for patterns 4 and 5, we

understand that the robot saves 33 seconds by sus-

pending the place action of A in the bookcase. By

comparing the results for patterns 4 and 6, we under-

stand that the robot saves 55 seconds by suspending

the move action from node 2 to node 5. By compar-

ing the results for patterns 4 and 7, we understand that

the robot saves 64 seconds by omitting the move ac-

tion from node 2 to node 5. From these results, we

understand that the robot saves tens of seconds by re-

planning and action/plan suspension. We also under-

stand that the sooner the robot replans, the shorter the

total plan execution time becomes. The robot saved

much time not only by skipping unnecessary pick-

and-place actions but also by omitting unnecessary

move actions.

Table 2: Time for Moving the Object to the Trash Can.

Pattern 4 5 6 7

Time(Sec) 143 110 88 79

10 CONCLUSIONS

In this paper, we have integrated planning, action

preparation, replanning, and suspension of action ex-

ecution/preparation. We have shown that this inte-

gration is effective not only for adapting to dynamic

environments but also for saving plan execution time

in manipulation. We implemented this integration by

the three strategies shown in this paper. We have

also shown that this integration method is effective

for pick-and-place manipulation in dynamic environ-

ments using a real mobile robot with two arms.

In future, we would like to explore more possi-

bilities of speculative action preparations, replanning,

and action/plan suspensions in other scenarios. When

the robot has nothing to do, the robot might be able to

prepare for future tasks. However, this is not simple

because the speculative action preparation might pre-

vent or delay future plan execution if the speculation

is wrong.

ACKNOWLEDGEMENTS

A part of this work is supported by a grant from the

Intelligent Robot Technology (RT) Software Project

of the New Energy and Industrial Technology Devel-

opment Organization (NEDO) of Japan. We would

ActionPreparationandReplanninginManipulation

115

like to thank the researchers of Yaskawa Corporation

who cooperated with us in conducting experiments.

REFERENCES

Ambros-Ingerson, J. and Steel, S. (1988). Integrating plan-

ning, execution and monitoring. In AAAI88, pages

735–740.

Cambon, S., Alami, R., and Gravot, F. (2009). A hybrid

approach to intricate motion, manipulation and task

planning. Journal of Robotics Research, 28(1):104–

126.

Chapman, D. (1987). Planning for conjunctive goals. Arti-

ﬁcial Intelligence, 32(3):333–377.

Choi, J. and Amir, E. (2009). Combining planning and mo-

tion planning. In ICRA09, pages 4374–4380.

Fikes, R., Hart, P., and Nilsson, N. (1972). Learning and

executing generalized robot plans. Artiﬁcial Intelli-

gence, 3(4):251–288.

Fikes, R. and Nilsson, N. (1971). STRIPS: A new approach

to the application of theorem proving to problem solv-

ing. Artiﬁcial Intelligence, 2(3-4):189–208.

Haspalamutgil, K., Palaz, C., Uras, T., Erdem, E., and

Patoglu, V. (2010). A tight integration of task plan-

ning and motion planning in an execution monitoring

framework. In AAAI10 Workshop on Bridging The

Gap Between Task And Motion Planning (BTAMP).

Hauser, K. and Latombe, J. (2009). Integrating task and

PRM motion planning: Dealing with many infeasible

motion planning queries. In ICAPS09 Workshop on

Bridging the Gap between Task and Motion Planning

(BTAMP).

Hayashi, H., Ogawa, H., and Matsuhira, N. (2013).

Htn planning for pick-and-place manipulation. In

ICAART13, pages 383–388.

Hayashi, H., Tokura, S., Hasegawa, T., and Ozaki, F.

(2006). Dynagent: An incremental forward-chaining

HTN planning agent in dynamic domains. In Declar-

ative Agent Languages and Technologies III, number

3904 in LNAI, pages 171–187. Springer.

Hayashi, H., Tokura, S., Ozaki, F., and Doi, M. (2009).

Background sensing control for planning agents work-

ing in the real world. International Journal of Intel-

ligent Information and Database Systems, 3(4):483–

501.

Jacobs, S. A., Manavi, K., Burgos, J., Denny, J., Thomas,

S., and Amato, N. M. (2012). A scalable method

for parallelizing sampling-based motion planning al-

gorithms. In ICRA12, pages 2529–2536.

Kaelbling, L. P. and Lozano-Perez, T. (2010). Hierarchi-

cal task and motion planning in the now. In ICRA10

Workshop on Mobile Manipulation.

Einig, L., Klimentjew, D., Rockel, S., Zhang, L., and

Zhang, J. (2013). Parallel plan execution and re-

planning on a mobile robot using state machines with

HTN Planning Systems. In ROBIO13, pages 151–157.

LaValle, S. M. and Kuffner, J. J. (2001). Rapidly-exploring

random trees: Progress and prospects, pages 293–

308. A K Peters.

Mainprice, J. and Berenson, D. (2014). Motion planning

for human-robot collaborative manipulation tasks us-

ing prediction of human motion. In RSS14 Workshop

on Human-Robot Collaboration for Industrial Manu-

facturing.

McAllester, D. and Rosenblitt, D. (1991). Systematic non-

linear planning. In AAAI91, pages 634–639.

Nau, D., Cao, Y., Lotem, A., and M˜unoz-Avila, H. (1999).

SHOP: simple hierarchical ordered planner. In IJ-

CAI99, pages 968–975.

Nau, D., Ilghami, O., Kuter, U., Murdock, W. J., Wu, D.,

and Yaman, F. (2003). SHOP2: an HTN planning

system. Journal of Artiﬁcial Intelligence Research,

20:379–404.

Nieuwenhuisen, D., van der Stappen, A. F., and Overmars,

M. H. (2008). An effective framework for path plan-

ning amidst movable obstacles. In Algorithmic Foun-

dation of Robotics VII, pages 87–102.

Ota, J. (2009). Rearrangement planning of multiple mov-

able objects by a mobile robot. Advanced Robotics,

23(1-2):1–18.

Park, C., Pan, J., and Manocha, D. (2013). Real-

time optimization-based planning in dynamic environ-

ments using GPUs. In ICRA13, pages 4090–4097.

Philip, M., Chestnutt, J., Kagami, S., Nishiwaki, K.,

Kuffner, J., and Kanabe, T. (2009). Humanoid nav-

igation planning using future perceptive capability. In

Digital Human Symposium.

Russell, S. and Norvig, P. (1995). Artiﬁcial Intelligence: A

Modern Approach. Prentice Hall.

Stilman, M., Schamburek, J.-U., Kuffner, J., and Asfour, T.

(2007). Manipulation planning among movable obsta-

cles. In ICRA07, pages 3327–3332.

Tate, A. (1977). Generating project networks. In IJCAI77,

pages 888–893.

Wolfe, J., Marthi, B., and Russel, S. (2010). Combined

task and motion planning for mobile manipulation. In

ICAPS10, pages 254–257.

Yaskawa Electric Corporation. (2007). YASKAWA devel-

ops a service robot “SmartPal V (SmartPal Five)”-

Additional waist area unit expands the allowable op-

erating range -, http://www.yaskawa.co.jp/en/topics/

071128

01/index.html, (Accessed on 10 June 2014).

Zacharias, F., Borst, C., and Hirzinger, G. (2006). Bridging

the gap between task planning and path planning. In

IROS06, pages 4490 – 4495.

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

116