26 Front. Comput. Sci.,2024,0(0): 1–42
code. Its features include providing detailed ex- abilities for planning, reasoning, and collaboration
planationsoferrormessages,suggestingpotential in embodied environments. In specific, [140] pro-
fixes,andensuringtheaccuracyofthecode. PEN- poses a unified agent system for embodied reason-
TESTGPT[125]isapenetrationtestingtoolbased ingandtaskplanning. Inthissystem,theauthorsde-
on LLMs, which can effectively identify common sign high-levelcommands to enable improvedplan-
vulnerabilities,andinterpretsourcecodetodevelop ningwhileproposelow-levelcontrollerstotranslate
exploits. DB-GPT [41] utilizes the capabilities of commandsintoactions. Additionally,onecanlever-
LLMstosystematicallyassesspotentialrootcauses agedialoguestogatherinformation[141]toaccel-
ofanomaliesindatabases. Throughtheimplementa- eratetheoptimizationprocess. [142,143]employ
tionofatreeofthoughtapproach,DB-GPTenables autonomousagentsforembodieddecision-making
LLMs to backtrack to previous steps in case the and exploration. To overcome the physical con-
current step proves unsuccessful, thus enhancing straints, the agents can generate executable plans
theaccuracyofthediagnosisprocess. andaccomplishlong-termtasksbyleveragingmulti-
pleskills. Interms ofcontrolpolicies,SayCan [78]
Industrial Automation: In the field of indus-
focusesoninvestigatingawiderangeofmanipula-
trialautomation,LLM-basedagentscanbeusedto
tionandnavigationskillsutilizingamobilemanip-
achieveintelligentplanningandcontrolofproduc-
ulator robot. Takinginspiration from typical tasks
tion processes. [129] proposes a novel framework
encountered in a kitchen environment, it presents
thatintegrateslargelanguagemodels(LLMs)with
acomprehensiveset of551skillsthat coverseven
digital twin systems to accommodate flexible pro-
skill families and 17 objects. These skills encom-
duction needs. The framework leverages prompt
passvariousactionssuchaspicking,placing,pour-
engineering techniques to create LLM agents that
ing,grasping,andmanipulatingobjects,amongoth-
can adapt to specific tasks based on the informa-
ers. TidyBot[137]isanembodiedagentdesigned
tion provided by digital twins. These agents can
topersonalizehouseholdcleanuptasks. Itcanlearn
coordinate a series of atomic functionalities and
skillsto completeproduction tasksat differentlev- users’preferencesonobjectplacementandmanipu-
lationmethodsthroughtextualexamples.
els within the automation pyramid. This research
demonstratesthepotentialofintegratingLLMsinto
To promote the application of LLM-based au-
industrialautomationsystems,providinginnovative
tonomousagents,researchershavealsointroduced
solutionsformoreagile,flexibleandadaptivepro-
manyopen-sourcelibraries,basedonwhichthede-
ductionprocesses. IELLM[130]showcasesacase
veloperscanquicklyimplementandevaluateagents
study on LLMs’ role in the oil and gas industry,
according to their customized requirements [19,81,
covering applications like rock physics, acoustic
127,144–157]. Forexample,LangChain[149]isan
reflectometry,andcoiledtubingcontrol.
open-sourceframeworkthatautomatescoding,test-
Robotics & Embodied Artificial Intelligence: ing,debugging,anddocumentationgenerationtasks.
Recent works have developed more efficient rein- Byintegratinglanguagemodelswithdatasources
forcementlearningagentsforroboticsandembod- and facilitating interaction with the environment,
ied artificial intelligence [16,38,78,132–135,140– LangChainenablesefficientandcost-effectivesoft-
143]. Thefocusisonenhancingautonomousagents’ ware development through natural language com-