Robot Task Planning and Situation Handling in Open Worlds

2025-04-15 0 0 2.87MB 7 页 10玖币

侵权投诉

Yan Ding1, Xiaohan Zhang1, Saeid Amiri1, Nieqing Cao1, Hao Yang2, Chad Esselink2, Shiqi Zhang1

Abstract— Automated task planning algorithms have been

developed to help robots complete complex tasks that require

multiple actions. Most of those algorithms have been developed

for “closed worlds” assuming complete world knowledge is pro-

vided. However, the real world is generally open, and the robots

frequently encounter unforeseen situations that can potentially

break the planner’s completeness. This paper introduces a

novel algorithm (COWP) for open-world task planning and

situation handling that dynamically augments the robot’s action

knowledge with task-oriented common sense. In particular,

common sense is extracted from Large Language Models based

on the current task at hand and robot skills. For systematic

evaluations, we collected a dataset that includes 561 execution-

time situations in a dining domain, where each situation

corresponds to a state instance of a robot being potentially

unable to complete a task using a solution that normally works.

Experimental results show that our approach signiﬁcantly

outperforms competitive baselines from the literature in the suc-

cess rate of service tasks. Additionally, we have demonstrated

COWP using a mobile manipulator. The project website is

available at: https://cowplanning.github.io/, where

a more detailed version can also be found. This version has

been accepted for publication in Autonomous Robots.

I. INTRODUCTION

Robots that operate in the real world frequently encounter

complex tasks that require multiple actions. Automated task

planning algorithms have been developed to help robots

sequence actions to accomplish those tasks [1]. Closed world

assumption (CWA) is a presumption that was developed

by the knowledge representation community and states that

“statements that are true are also known to be true” [2]. Most

current task planners have been developed for closed worlds,

assuming complete domain knowledge is provided and one

can enumerate all possible world states [3]–[6]. However,

the real world is “open” by nature, and unforeseen situations

are common in practice [7]. As a consequence, current

automated task planners tend to be fragile in open worlds rife

with situations. Fig. 1 shows an example situation: Aiming

to grasp a cup for drinking water, a robot found the cup

was occupied with forks and knives. Although one can name

many such situations, it is impossible to provide a complete

list of them. To this end, researchers have developed open-

world planning methods for robust task completions in real-

world scenarios [7]–[12].

There are mainly two ways of performing open-world

planning without humans in the loop in the literature. One

relies on dynamically building an external knowledge base to

1Yan Ding, Xiaohan Zhang, Saeid Amiri, Nieqing Cao, and Shiqi Zhang

are with SUNY Binghamton. {yding25, xzhan244, samiri1,

ncao1, zhangs}@binghamton.edu

2Hao Yang and Chad Esselink are with Ford Motor Company.

{hyang1, cesselin}@ford.com

Wrist Camera

Gripper

UR5e Arm

Segway Base

CPU

Battery

Fig. 1. An illustrative example of a situation in the real world, encountered

during the execution of the plan “delivering a cup to a human for drinking

water.” The robot approached a cabinet in a kitchen room on which a cup

was located. The robot then found the cup to be delivered. Before grasping it,

however, the robot detected a situation that the cup was occupied with a fork,

a knife, and a spoon. This situation prevented the robot from performing

the current action (i.e., grasping) and rendered normal solutions for drinking

water invalid. A mobile service robot with a UR5e arm on a Segway RMP-

110 base was used for demonstrations in this work.

assist a pre-deﬁned task planner, where this knowledge base

is usually constructed in an automatic way using external

information [7]–[9]. Such external knowledge bases are con-

sidered bounded due to their representation and knowledge

source, which limits the “openness” of their task planners.

Another (more recent) way of building open-world planners

is to leverage Large Language Models (LLMs) [13]. LLMs

have signiﬁcantly improved the performance of downstream

language tasks in recent years [14]–[17]. Recent research

has demonstrated that those LLMs contain a wealth of

commonsense knowledge [12], [18]–[20]. While it is an in-

tuitive idea of extracting common sense from LLMs for task

planning [10]–[12], [21], there is a fundamental challenge

for robots to “ground” domain-independent commonsense

knowledge [22] to speciﬁc domains that are featured with

many domain-dependent constraints. We propose to acquire

common sense from LLMs, and aim to improve the task

completion and situation handling skills of service robots.

In this paper, we develop a robot task planning framework,

called Common sense-based Open-World Planning (COWP),

that uses an LLM (GPT-3 in our case [14]) for dynam-

ically augmenting automated task planners with external

task-oriented common sense. COWP is based on classical

planning and leverages LLMs to augment action knowledge

(action preconditions and effects) for task planning and

situation handling. The main contribution of this work is

a novel integration of a pre-trained LLM with a knowledge-

based task planner. Inheriting the desirable features from

both sides, COWP is well grounded in speciﬁc domains while

arXiv:2210.01287v2 [cs.RO] 29 Sep 2024

embracing commonsense solutions at large.

For systematic evaluations, we have created a dataset

with 561 execution-time situations collected from a dining

domain [23]–[25] using a crowd-sourcing platform, where

each situation corresponds to an instance of a robot not being

able to perform a plan (that normally works). According to

experimental results, we see COWP performed signiﬁcantly

better than three literature-selected baselines [6], [8], [11]

in success rate. We implemented and demonstrated COWP

using a mobile manipulator.

II. BACKGROUND AND RELATED WORK

In this section, we ﬁrst brieﬂy discuss classical task

planning methods that are mostly developed under the closed

world assumption. We then summarize three families of

open-world task planning methods for robots, which are

grouped based on how unforeseen situations are addressed.

Classical Task Planning for Closed Worlds: Closed world

assumption (CWA) indicates that an agent is provided with

complete domain knowledge, and that all statements that are

true are known to be true by the agent [2]. Most automated

task planners have been developed under CWA [1], [5],

[26]. Although robots face the real world that is open by

nature, their planning systems are frequently constructed

under the CWA [7], [27], [28]. The consequence is that

those robot planning systems are not robust to unforeseen

situations at execution time. In this paper, we aim to develop

a task planner that is aware of and able to handle unforeseen

situations in open-world scenarios.

Open-World Task Planning with Human in the Loop:

Task planning systems have been developed to acquire

knowledge via human-robot interaction to handle open-

world situations [29]–[31]. For instance, researchers created

a planning system that uses dialog systems to augment

their knowledge bases [29], whereas Amiri et al. (2019)

further modeled the noise in language understanding [30].

Tucker et al. (2020) enabled a mobile robot to ground new

concepts using visual-linguistic observations, e.g., to ground

the new word “box” given command of “move to the box”

by exploring the environment and hypothesizing potential

new objects from natural language [31]. The major difference

from those open-world planning methods is that COWP does

not require human involvement.

Open-World Task Planning with External Knowledge:

Some existing planning systems address unforeseen situa-

tions by dynamically constructing an external knowledge

base for open-world reasoning. For instance, researchers have

developed object-centric planning algorithms that maintain a

database about objects and introduce new object concepts

and their properties (e.g., location) into their task plan-

ners [8], [9]. For example, Jiang et al. (2019) developed an

object-centric, open-world planning system that dynamically

introduces new object concepts through augmenting a local

knowledge base with external information [8]. In the work

of Hanheide et al. (2017), additional action effects and

assumptive actions were modeled as an external knowledge

to explain the failure of task completion and compute plans

in open worlds [7]. A major difference from their methods is

that COWP employs an LLM that is capable of responding

to any situation, whereas the external knowledge sources of

those methods limits the openness of their systems.

Open-World Task Planning with LLMs: LLMs can encode

a large amount of common sense from corpus [32] and

have been applied to robot systems to complete high-level

tasks. Example LLMs include BERT [33], GPT-2 [17], GPT-

3 [14], and OPT [15]. For example, Kant et al. (2022)

developed a household robot that uses a ﬁne-tuned LLM

to reason about rearrangements of new objects [10]. Other

teams used LLMs to compute plans for high-level tasks

speciﬁed in natural language (e.g., “make breakfast”) by

sequencing actions [11], [12], [34]. Different from the above-

mentioned approaches, in addition to commonsense knowl-

edge extracted from LLMs, our system utilizes rule-based

action knowledge from human experts. As a result, our

planning system can be better grounded to speciﬁc domains,

and is able to incorporate common sense to augment robot

capabilities supported by predeﬁned skills.

III. ALGORITHM

In this section, we ﬁrst provide a problem statement

and then present our open-world planning approach called

Common sense-based Open-World Planning (COWP).

A. Problem Description

A classical task planning problem is deﬁned by a do-

main description and a problem description. The domain

description includes a set of actions, and action preconditions

and effects. The problem description includes an initial state

and goal conditions. In this paper, such a classical planning

system is referred to as a closed-world task planner. In our

case, we provide the robot with a predeﬁned closed-world

task planner (implemented using PDDL [35]), and an LLM

(GPT-3 in our case). PDDL, an action-centered language,

is designed to formalize Artiﬁcial Intelligence (AI) planning

problems, allowing for a more direct comparison of planning

algorithms and implementations [35]. A situation is deﬁned

as an unforeseen world state that potentially prevents an

agent from completing a task using a solution that normally

works. The goal of an open-world planner is to compute

plans and handle situations towards completing service tasks

or reporting “no solution” as appropriate.

B. Algorithm Description

Fig. 2 illustrates the three major components (yellow

boxes) of our COWP framework. Task Planner is used

for computing a plan under the closed-world assumption

and is provided as prior knowledge in this work. Plan

Monitor evaluates the overall feasibility of the current plan

using common sense. Knowledge Acquirer is for acquiring

common sense to augment the robot’s action effects when

the task planner generates no plan.

Algorithm 1 describes how the components of COWP

interact with each other. Initially, Task Planner generates a

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：

RobotTaskPlanningandSituationHandlinginOpenWorldsYanDing1,XiaohanZhang1,SaeidAmiri1,NieqingCao1,HaoYang2,ChadEsselink2,ShiqiZhang1Abstract—Automatedtaskplanningalgorithmshavebeendevelopedtohelprobotscompletecomplextasksthatrequiremultipleactions.Mostofthosealgorithmshavebeendevelopedfor“closedworlds...

展开>> 收起<<

Robot Task Planning and Situation Handling in Open Worlds.pdf

共7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！

Robot Task Planning and Situation Handling in Open Worlds

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: