Learning Action Duration and Synergy in Task Planning for Human-Robot Collaboration Samuele Sandrini1 Marco Faroni1 Nicola Pedrocchi1

2025-05-02 0 0 2.2MB 6 页 10玖币

侵权投诉

Learning Action Duration and Synergy in Task Planning for

Human-Robot Collaboration

Samuele Sandrini1, Marco Faroni1, Nicola Pedrocchi1

Abstract—A good estimation of the actions’ cost is key in

task planning for human-robot collaboration. The duration of

an action depends on agents’ capabilities and the correlation

between actions performed simultaneously by the human and

the robot. This paper proposes an approach to learning actions’

costs and coupling between actions executed concurrently by

humans and robots. We leverage the information from past

executions to learn the average duration of each action and a

synergy coefﬁcient representing the effect of an action performed

by the human on the duration of the action performed by the

robot (and vice versa). We implement the proposed method in

a simulated scenario where both agents can access the same

area simultaneously. Safety measures require the robot to slow

down when the human is close, denoting a bad synergy of tasks

operating in the same area. We show that our approach can

learn such bad couplings so that a task planner can leverage this

information to ﬁnd better plans.

Index Terms—Human-Robot Interaction; Task And Motion

Planning; Task planning; Learning for task planning.

I. INTRODUCTION

Human-Robot Collaboration (HRC) often requires the system

to make decisions based on human users’ observed or predicted

behavior. In such a context, planning and allocating tasks to the

human and robot agents are two complex problems, even for

tasks composed of a few activities. Two modelling issues are

essential. First, human behavior is intrinsically unpredictable

and partially uncontrollable [

]. Second, the coupling between

the human and the robotic agents is characterized by a large

variability. For example, the feasibility and the duration of

an action may vary because of the interference between the

human and the robot (e.g., safety stops of the robot or even

due to path-switch in motion re-planning [2]).

In recent years, task planning and allocation problems for

HRC have been investigated. Existing works tried to model

human preferences explicitly and ergonomics into planning

[

]. For example, [

], [

] proposed a hierarchical agent-based

task planner, where complex tasks can be decomposed into

simpler actions. This approach can improve the collaborative

experience by considering human preferences as social costs [

[

]. Manufacturing-oriented works focus on process throughput

by minimizing the expected duration [

] or planning contingent

plans to reduce process errors [

], [

]. A promising approach

to deal with uncertainty on the task duration is timeline-based

planning [

]. Timeline-based planning explicitly models the

duration variability of tasks and ﬁnds plans that are robust with

This work was partially supported by ShareWork project (H2020, European

Commission – G.A. 820807).

STIIMA-CNR - Institute of Intelligent Industrial Tech-

nologies and Systems, National Research Council of Italy

{name.surname}@stiima.cnr.it

respect to it. Timeline-based planning was used in HRC with

pre-computed robot motions [

], and online motion planning

[

], demonstrating robust plans allow for less frequent re-

planning and shorter average task duration.

Most planning-based methods assume a guess of the action

cost (e.g., the duration) is available from a domain expert.

If this information is unreliable, the task planner reasons on

wrong assumptions so that the resulting plans become sub-

optimal or even infeasible (leading to frequent re-planning

when using a motion re-planner such as [

]). Moreover, the

duration guess does not consider possible couplings between

the tasks executed concurrently by the human and the robot.

Therefore, the task planner neglects each agent’s positive or

negative effects on the others. For example, if concurrent tasks

require the human and the robot to work in the same area, the

robot would probably slow down because of safety. If the task

planner knew this effect, it could favour pairs of tasks that

avoid robot safety slowdowns.

In this work, we propose a method to learn the expected

duration of a joint plan. We leverage a minimum-time formu-

lation of the task planning/allocation problem for HRC. Then,

we estimate the expected duration of each task from previous

executions and learn a synergy coefﬁcient that represents the

effect of one task over the other. A synergy coefﬁcient greater

than one means that the human task causes a slow down of the

robot task. We show that it is possible to cast the estimation

of the synergy coefﬁcients into a set of linear regression

problems (one for each task). The resulting synergy coefﬁcients

can be exploited in task planning and allocation method to

select advantageous task couplings. The proposed approach

can be used ofﬂine – to obtain a guess of the task duration

and synergies – or iteratively as the number of executions

grows to reﬁne the initial guess over time. We demonstrate the

proposed approach in an HRC scenario where a human and a

collaborative robot shall perform a sequence of pick-and-place

operations. Experiments show that pairs of tasks that drive the

nearby agents lead to high synergy coefﬁcients, recognizing

favorable and unfavourable pairs.

The paper is organized as follows. Section II deﬁnes the

task planning problem for HRC systems. Section III builds on

that model to formulate the task expected duration in terms of

average duration and synergy coefﬁcients. Then, it casts the

estimation of such coefﬁcients into a set of linear regression

problems. Section IV describes the software architecture to

collect and process the data from task executions. Section

V applies the proposed methodology to the manufacturing

example and shows that the resulting coefﬁcients reﬂect the

good and bad coupling effects of robot safety stops. Finally,

arXiv:2210.11660v1 [cs.RO] 21 Oct 2022

conclusions and future works are discussed in Section VI.

II. PRELIMINARIES

A. Task Planning Problem

A task planning problem can be formalized as an optimiza-

tion problem. Given a set

{H, R}

of agents, i.e., human and

robot, and a set of Tasks

T={τi}

, the objective of the problem

is to obtain a task plan and assignment

that minimizes the

duration of the process.

We denote a task by a tuple τ= (l, d, t), in which:

•l∈ {H, R}

is an assignment variable that speciﬁes which

agent can perform the speciﬁc task, i.e., human, or robot;

•d∈R+is a guess of the task duration;

•t=tstart ;tend

is an interval with endpoints correspond-

ing to the start and end time.

We refer to

and

as the subsets of

such that

l=H

and l=R, respectively.

In this context, it is possible to introduce a binary assignment

variable that deﬁnes the allocation of

τi∈ T

to the robot (

)

or to the human (aH

i):

i=(1,if task τiis assigned to the robot

0,otherwise

i=(1,if task τiis assigned to the human

0,otherwise.

(1)

The duration of a plan

is the maximum between the duration

of the robot’s and the human’s plan, denoted by

and

respectively. The duration of each agent’s plan can be calculated

as:

π=X

iaH

π=X

iaR

(2)

By deﬁning a cost function

that represents the duration of a

plan πas:

J= max{dH

π, dR

π}(3)

the optimal plan π∗is:

π∗= argmin

J(4)

subject to the constraint that

τi∈ T

can only be assigned

during π:

i+aR

i= 1 ∀is.t. τi∈ T (5)

and assignment, temporal, and causal constraints owed to the

process requirements1.

(4)

only involves assignment constraints, the problem is a task allocation;

if causal and/or temporal constraints are considered, the problem becomes a

task planning&scheduling problem.

III. METHODOLOGY

The task planning problem assumes that a guess of the task

duration exists. A domain expert usually provides this guess.

However, this information is often unreliable and comes with a

sizeable unmodeled uncertainty. Moreover, it does not account

for the effect of the task performed by other agents. Indeed, the

task duration is a function of the tasks executed simultaneously

by the other agent, i.e.:

i=dR

i(τH

j)∀js.t. τH

j:tH

j∩tR

i6=∅

i=dH

i(τR

j)∀js.t. τR

j:tR

j∩tH

i6=∅

where

τH

and

τR

are tasks assigned to humans and robots. If

the coupling between tasks is not modelled, the optimal task

(4)

may be unreliable when executed on the real-world system.

To overcome this problem, we extend the formulation given in

II-A to include a synergy coefﬁcient in the task duration.

A. Task Planning Problem with Explicit Task Coupling

For each couple of task indices

(i, j)

, a synergy term is

introduced for each agent and denoted with

i,j

for the robot

agent and

i,j

for the human agent. The synergy term denotes

the increment of the duration of task

τi

when executed by the

robot while the human is executing task

τj

. We deﬁne this

coefﬁcient as:

i,j =dR

i,j

(6)

where

i,j

is the expected duration of task

τi

when the human

executes

τj

and

is the expected value of the duration of

τR

for all concurrent tasks

τH

. Thus, the duration of a plan

becomes:

π=X

iaR

iX

i,j δR

i,j aH

j(7)

where

δR

i,j

represents the ratio of the overlapping time between

τR

and

τH

with respect to the duration of the task

τR

, thus

deﬁned as:

δR

i,j =D(tH

j∩tR

D(tR

i)(8)

and

is a function that calculates the duration of a temporal

interval t= [tstart, tend]:

D(t) = (tend −tstart,if t6=∅

0,if t=∅(9)

Equations

(6)

(7)

(8)

can be rewritten for the human agent

as:

i,j =dH

i,j

(10)

δH

i,j =D(tR

j∩tH

D(tH

i)(11)

π=X

iaH

iX

i,j δH

i,j aR

j(12)

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

LearningActionDurationandSynergyinTaskPlanningforHuman-RobotCollaborationSamueleSandrini1,MarcoFaroni1,NicolaPedrocchi1AbstractAgoodestimationoftheactions'costiskeyintaskplanningforhuman-robotcollaboration.Thedurationofanactiondependsonagents'capabilitiesandthecorrelationbetweenactionsperformedsimu...

展开>> 收起<<

Learning Action Duration and Synergy in Task Planning for Human-Robot Collaboration Samuele Sandrini1 Marco Faroni1 Nicola Pedrocchi1.pdf

共6页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Learning Action Duration and Synergy in Task Planning for Human-Robot Collaboration Samuele Sandrini1 Marco Faroni1 Nicola Pedrocchi1

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: