Embodied AI in Mobile Robot Simulation with EyeSim: Coverage Path

Planning with Large Language Models

Xiangrui Kong

1 a

, Wenxiao Zhang

2 b

, Jin Hong

2 c

and Thomas Br

aunl

1 d

Department of Electrical, Electronic and Computer Engineering, University of Western Australia, Crawley, Australia

Department of Computer Science and Software Engineering, University of Western Australia, Crawley, Australia

Keywords:

Natural Language Processing, Mobile Robots, Path Planning, Indoor Navigation.

Abstract:

In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities in understanding

and solving mathematical problems, leading to advancements in various ﬁelds. We propose an LLM-embodied

path planning framework for mobile agents, focusing on solving high-level coverage path planning issues and

low-level control. Our proposed multi-layer architecture uses prompted LLMs in the path planning phase and

integrates them with the mobile agents’ low-level actuators. To evaluate the performance of various LLMs,

we propose a coverage-weighted path planning metric to assess the performance of the embodied models. Our

experiments show that the proposed framework improves LLMs’ spatial inference abilities. We demonstrate

that the proposed multi-layer framework signiﬁcantly enhances the efﬁciency and accuracy of these tasks by

leveraging the natural language understanding and generative capabilities of LLMs. Experiments conducted

in our EyeSim simulation demonstrate that this framework enhances LLMs’ 2D plane reasoning abilities and

enables the completion of coverage path planning tasks. We also tested three LLM kernels: gpt-4o, gemini-

1.5-ﬂash, and claude-3.5-sonnet. The experimental results show that claude-3.5 can complete the coverage

planning task in different scenarios, and its indicators are better than those of the other models. We have made

our experimental simulation platform, EyeSim, freely available at https://roblab.org/eyesim/.

1 INTRODUCTION

The application of Large Language Models (LLMs)

has grown exponentially, revolutionizing various

ﬁelds with their advanced capabilities (Hadi et al.,

2023). Modern LLMs have evolved to perform var-

ious tasks beyond natural language processing. When

integrated into mobile agents, these LLMs can inter-

act with the environment and perform tasks without

the need for explicitly coded policies or additional

model training. This capability leverages the exten-

sive pre-training of LLMs, enabling them to general-

ize across tasks and adapt to new situations based on

their understanding of natural language instructions

and contextual cues.

Embodied AI refers to artiﬁcial intelligence sys-

tems integrated into physical entities, such as mo-

bile robots, that interact with the environment through

https://orcid.org/0000-0001-5066-1294

https://orcid.org/0009-0000-5196-8562

https://orcid.org/0000-0003-1359-3813

https://orcid.org/0000-0003-3215-0161

sensors and actuators (Chrisley, 2003). The integra-

tion of LLMs with embodied AI in applications such

as autonomous driving (Dorbala et al., 2024) and hu-

manoid robots (Cao, 2024) demonstrates their poten-

tial. However, the application of LLMs in controlling

mobile robots remains challenging due to issues such

as end-to-end control gaps, hallucinations, and path

planning inefﬁciencies. LLMs possess the capability

to solve mathematical problems, which directly aids

in path planning methods (Gu, 2023).

Path planning and obstacle avoidance are critical

for the effective operation of mobile robots, ensur-

ing safe and efﬁcient navigation in dynamic environ-

ments (Hewawasam et al., 2022). Coverage path plan-

ning is a typical method employed in various research

areas, such as ocean seabed mapping (Galceran and

Carreras, 2012), terrain reconstruction (Torres et al.,

2016), and lawn mowing (Hazem et al., 2021). Tra-

ditional path planning methods include algorithms

such as A* (Warren, 1993), D* (Ferguson and Stentz,

2005), and potential ﬁeld methods (Barraquand et al.,

1992). Given a global map, a path-planning method

can be framed as a mathematical problem solvable by

Kong, X., Zhang, W., Hong, J. and Bräunl, T.

Embodied AI in Mobile Robot Simulation with EyeSim: Coverage Path Planning with Large Language Models.

DOI: 10.5220/0013455100003970

In Proceedings of the 15th International Conference on Simulation and Modeling Methodologies, Technologies and Applications (SIMULTECH 2025), pages 185-192

ISBN: 978-989-758-759-7; ISSN: 2184-2841

185

(a) Standard lawnmower (b) Square spiral

Figure 1: Comparison of Path Planning Patterns generated

by prompted LLMs.

LLMs. In this context, we simplify some traditional

path-planning methods and test LLMs in our mobile

robot simulator. LLMs demonstrate their ability to

solve mathematical problems collaboratively (Zhang

et al., 2024). The EyeSim VR is a multiple mo-

bile robot simulator with VR functionality based on

game engine Unity 3D that allows experiments with

the same unchanged EyeBot programs that run on the

real robots (Br

aunl, 2020), which is capable of simu-

lating all major functionalities in RoBIOS-7. We have

made our experimental simulation platform, EyeSim,

freely available at https://roblab.org/eyesim/.

This paper presents a multi-layer coverage path

planner based on existing multimodal large language

models. It involves the static low-dimensional decon-

struction of unstructured maps, abstracting spatial re-

lationships into mathematical problems for reasoning

and solving by prompted LLMs. The reasoning accu-

racy of the LLM is enhanced through multi-turn dia-

logues and multimodal interactions. The inferred re-

sults from the LLM are combined with the control in-

terface, enabling the mobile agent to control the robot

in real time for path planning. Simulation experi-

ments demonstrate that LLMs possess path-planning

capabilities in unstructured static maps.

2 RELATED WORKS

2.1 LLMs in Mobile Robots

Currently, LLMs are involved in various aspects of

mobile robots, including code writing, model train-

ing, action interpretation, and task planning. LLMs

can process new commands and autonomously re-

compose API calls to generate new policy code

by chaining classic logic structures and referencing

third-party libraries (Liang et al., 2023). LLMs have

also been used to automatically generate reward al-

gorithms for training robots to learn tasks such as

pen spinning (Ma et al., 2024). PaLM-E, an embod-

ied language model trained on multi-modal sentences

combining visual, state estimation, and textual input

encodings, demonstrates the versatility and positive

transfer across diverse embodied reasoning tasks, ob-

servation modalities, and embodiments (Driess et al.,

2023). LLMs have shown promise in processing and

analyzing massive datasets, enabling them to uncover

patterns, forecast future occurrences, and identify ab-

normal behaviour in a wide range of ﬁelds (Su et al.,

2024). VELMA is an embodied LLM agent that gen-

erates the next action based on a contextual prompt

consisting of a verbalized trajectory and visual obser-

vations of the environment (Schumann et al., 2024).

Sharma et al. propose a method for using natural lan-

guage sentences to transform cost functions, enabling

users to correct goals, update robot motions, and re-

cover from planning errors, demonstrating high suc-

cess rates in simulated and real-world environments

(Sharma et al., 2022).

There is also some research applying LLMs in

zero-shot path planning (Chen et al., 2025). The

3P-LLM framework highlights the superiority of the

GPT-3.5-turbo chatbot in providing real-time, adap-

tive, and accurate path-planning algorithms compared

to state-of-the-art methods like Rapidly Exploring

Random Tree (RRT) and A* in various simulated sce-

narios (Latif, 2024). Singh et al. describe a program-

matic LLM prompt structure that enables the genera-

tion of plans functional across different situated envi-

ronments, robot capabilities, and tasks (Singh et al.,

2022). Luo et al. demonstrate the integration of a

sampling-based planner, RRT, with a deep network

structured according to the parse of a complex com-

mand, enabling robots to learn to follow natural lan-

guage commands in a continuous conﬁguration space

(Kuo et al., 2020). ReAct utilizes LLMs to generate

interleaved reasoning traces and task-speciﬁc actions

(Yao et al., 2022). These methods typically use LLMs

to replace certain components of mobile robots. The

development of a hot-swapping path-planning frame-

work centred around LLMs is still in its early stages.

2.2 Path Planning Method

Path planning for mobile robots involves determin-

ing a path from a starting point to a destination on

a known static map (Ab Wahab et al., 2024). Ob-

stacle avoidance acts as a protective mechanism for

SIMULTECH 2025 - 15th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

186

the robot, enabling interaction with obstacles encoun-

tered during movement. Low-level control connects

algorithms to different types of system agents, such

as UAVs, UGVs, or UUVs (Ali and Israr, 2024).

In addition to the A* and D* algorithms mentioned

in the previous chapter, path planning algorithms in-

clude heuristic optimization methods based on pre-

trained weights, such as genetic algorithms (Castillo

et al., 2007), particle swarm optimization (Dewang

et al., 2018), and deep reinforcement learning (Panov

et al., 2018). These pre-trained methods do not di-

rectly rely on prior knowledge but utilize data to pre-

train weights. The obstacle avoidance problem ad-

dresses dynamic obstacles encountered during move-

ment, ensuring the safety of the mobile agent. Main-

stream methods include the Artiﬁcial Potential Field.

The coverage path planning problem is a branch

of path planning problems. Compared with point-to-

point path planning, a coverage waypoint list needs to

cover the given area as much as possible (Di Franco

and Buttazzo, 2016). Classically, decomposing a

given map based on topological rules and then apply-

ing a repeatable coverage pattern is a common way to

solve this issue following the divide-and-conquer al-

gorithm (Palacios-Gas

os et al., 2018; Petitjean, 2002;

Smith, 1985). In this way, a known map is required to

start, whereas the Traveling Salesman Problem (TSP),

an optimization problem that seeks to determine the

shortest possible route for a salesman to visit a given

set of cities exactly once and return to the original

city, offers another solution to solve it in a node graph

(Hoffman et al., 2013). Figure 1 presented showcases

a comparison of four distinct path-planning patterns

employed in robotic navigation. The ﬁrst pattern, la-

belled as a standard lawnmower (Figure 1a), utilizes

a standard back-and-forth sweeping motion to ensure

comprehensive coverage of the area. The second pat-

tern, square spiral (Figure 1b), depicts a robot fol-

lowing an inward spiral trajectory, efﬁciently cover-

ing the space in a continuous inward motion. The

third pattern, square move (Figure 1c), illustrates a

robot navigating in a sequential inward square for-

mation, progressively moving towards the centre. Fi-

nally, the lawnmower after wall following (Figure 1d)

combines two approaches: initially, the robot adheres

to the perimeter of the wall following area, and subse-

quently, it adopts a lawnmower pattern to cover the re-

maining interior space. This comparative analysis of

path planning strategies highlights the versatility and

application-speciﬁc advantages of each method in en-

suring thorough area coverage in robotic navigation

tasks.

3 METHODOLOGY

As depicted in Figure 2, our method is divided into

three main sections: global planning, waypoint eval-

uation, and navigation. In global planning phase, a

coverage planning task in a given map is decomposed

into a cell map, and the additional requirement is de-

signed using natural language with a simpliﬁed for-

mat to decompose LLM responses. During the way-

point evaluation phase, the LLM responses are further

evaluated before execution. The theoretical coverage

rate and the theoretical shortest path distance are cal-

culated in this phase. Once the desired path passes

the evaluation, the planned warpaint list transitions to

the navigation phase. In navigation phase, the mo-

bile agent simply travels through them one by one and

triggers the safety mechanism if the sensor shows a

threshold distance between the robot and an unknown

obstacle.

3.1 Global Planning

We design a waypoint generation prompt with natural

language describing 2D grid maps like a chessboard

to simplify the inference difﬁculty of LLMs. During

the global phase, a prompt contains the size of the

grid map, current location, and response format. We

assume the LLM generates the desired waypoint list

with a required format which is a local position se-

quence separated with a bar sign. In order to evaluate

the performance and excitability of the planned path,

the desired waypoint list is visualised and calculated

in the phrase of waypoint evaluation. Considering the

robot’s kinematic limitation, we prompt a description

of mobile agents including equipped sensors, driving

commands, and basic status. We experimented with

various settings to describe robot behaviors in con-

versations with ChatGPT. However, we observed that

these changes in description had minimal impact on

the output responses. We use OpenAI GPT-4o ser-

vices (Achiam et al., 2023), a multimodal efﬁcient

model for inference and reasoning. The temperature

parameter with the range from 0 to 2 is set as 0.6 with

our prompt for a consistent planned path. Lower val-

ues for temperature result in more consistent outputs,

while higher values generate more diverse and cre-

ative results.

3.2 Waypoint Evaluation

The response from the LLMs can occasionally be in-

correct, leading us to design a waypoint evaluator to

mitigate hallucinations. Initially, the desired way-

point list is visualized on a 2D map, providing a clear

Embodied AI in Mobile Robot Simulation with EyeSim: Coverage Path Planning with Large Language Models

187

Figure 2: Multi-layer embodied path planning framework.

and precise layout of the proposed route. The shortest

path and the number of turns are then calculated math-

ematically to ensure efﬁciency and feasibility. Paths

that do not meet the required criteria are rejected and

not converted into a driving command list. The de-

signed dialogue system initiates as soon as the agent

receives the task command and map, continuing un-

til a waypoint list passes the evaluation. This ensures

that only optimal routes are considered for execution.

Once the mobile agent begins driving, the task cannot

be altered, guaranteeing consistency and reliability in

task completion.

Algorithm 1 begins by initializing key parameters:

the maximum number of iterations N, the evaluation

Data: N, θ, p

, s

Result: P ← {p

, s

}

initialization;

while n < N do

W ← Φ(P ) // LLM inference ;

r, τ ← E (W);

if r, τ > θ then

end

Algorithm 1: Initialization.

threshold θ, the target position p

, and the starting po-

sition s

. A prompt P is created, containing the task

description and current position, which is then used

by the LLM to generate waypoints. The LLM infer-

ence function Φ produces a list of waypoints W based

on this prompt, taking into account the grid map, cur-

rent location, and required response format. As the al-

gorithm iterates, it evaluates the generated waypoints

using the evaluation function E , which calculates the

shortest path r and the number of turns τ. If the cal-

culated path metrics r and τ exceed the predeﬁned

threshold θ, the waypoint list is considered feasible

and returned. This loop continues until a valid way-

point list is identiﬁed or the maximum number of it-

erations is reached. The algorithm ensures that only

optimal routes are considered, thus providing a ro-

bust framework for waypoint generation and evalu-

ation. This process incorporates global planning and

rigorous waypoint evaluation to leverage LLM capa-

bilities while ensuring safe and reliable path execu-

tion for mobile agents.

3.3 Waypoint Navigation

After evaluating the waypoint list, the mobile agent

begins to iterate through the waypoints. Due to po-

SIMULTECH 2025 - 15th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

188

while w

in W do

s ← O ;

′

← W ;

a ← Γ(s, s

′

) //Choose a method ;

∆ ← ||s − s

′

|| ;

if ∆ < d then

continue

end

Algorithm 2: Execution.

tential sensor errors and the intricacies of the path-

following method, it is essential for the mobile agent

to appropriately select the following method. Simple

waypoint following methods such as the dog curve

and turn-and-drive can be employed to navigate the

waypoints with a ﬁxed distance. These methods en-

able the mobile agent to follow the sequence of way-

points with smooth and accurate navigation along the

route.

In our approach, we decompose this procedure

using a status transform matrix that maps the next

driving command based on the current heading, cur-

rent position, and the next waypoint. This matrix al-

lows for dynamic adjustment and precise control dur-

ing navigation. Additionally, the designed safety sys-

tem ensures the execution is safe by preventing col-

lisions with unknown obstacles. This is achieved us-

ing a position-sensitive detector and LIDAR beams,

which continuously monitor the environment and pro-

vide real-time feedback for obstacle avoidance.

A algorithm 2 iterates over each waypoint w

the list W . The current position s is updated us-

ing odometry data O, and the next waypoint s

′

converted from the waypoint list W . The following

method is chosen based on the action command a,

which is determined by the selected path following

function Γ(s, s

′

). The distance ∆ between the current

position s and the next waypoint s

′

is calculated. If

the distance ∆ is less than a predeﬁned threshold d,

the algorithm continues to the next waypoint.

4 EXPERIMENT

4.1 Implement Details

This framework has been implemented on EyeBot

simulator (Br

aunl, 2023). The EyeBot simulator with

virtual reality EyeSim VR is a multiple mobile robot

simulator with VR functionality based on game en-

gine Unity 3D that allows experiments with the same

unchanged EyeBot programs that run on the real

robots. We adjust the environmental values based on

the task map from 5 × 5 to 11 × 11. In each map,

the mobile agent is at a random starting position and

runs the proposed method in 10 episodes, and all per-

formance metrics are averaged. Three large language

models are evaluated in the experiment including gpt-

4o, gemini-1.5-ﬂash and claude-3.5-sonnet with the

same system prompt and default temperature shown

in Figure 2.

4.2 Metrics

We referenced the metrics from (Anderson et al.,

2018) and (Zhao et al., 2021), including success rate,

average distance, and coverage rate. The success rate

indicates whether the paths generated by LLMs can

cover the designated area. Average distance repre-

sents the average path length of the mobile robot,

while coverage rate is a metric speciﬁc to coverage

methods, used to assess the completeness of coverage

path planning algorithms.

In traditional navigation evaluation standards, task

termination is determined by the distance between the

agent and the target point, which is effective for path

planning problems with clearly deﬁned start and end

points. However, for coverage path planning algo-

rithms, the generated paths do not have a clear end-

point, and the coverage path is autonomously decided

by the LLM. Therefore, we have added a coverage

rate metric to the comprehensive evaluation standards

referenced from the cited sources. Inspired by Suc-

cess weighted Path Length (SPL) from (Anderson

et al., 2018), we will refer to the following measure

as CPL, short for Coverage weighted by (normalized

inverse) Path Length:

CPL =

∑

i=1

max(p

, l

)

(1)

where N means the number of test episodes. A

and

indicate the area of the coverage path and the

area of the mission area, respectively. The ratio of A

and

is expressed as the Coverage Rate (CR), which

is used to evaluate the completeness of the path. The l

means the theoretical shortest path distance from the

mobile agent start point, and the p

is the Path Length

(PL) of the moving path by the agent.

4.3 Results and Analysis

The performance and time analysis are shown in Ta-

ble 1 and Table 2. All three models demonstrate

the ability to plan a coverage path in a square space

with a random start position. However, as the map

size increases, the coverage rate decreases by ap-

proximately 5% to 10%, though all models maintain

Embodied AI in Mobile Robot Simulation with EyeSim: Coverage Path Planning with Large Language Models

189

Table 1: Zero-shot coverage path planning performance using multiple LLM services in various environments.

Map Size

GPT-4o Gemini-1.5 Claude-3.5

CPL↑ PL↓ CR↓ CPL PL CR CPL PL CR

5 × 5 0.95 34.2 96.4 0.87 44.5 87.8 0.99 37.2 100

7 × 7 0.86 56.9 86.7 0.81 61.1 82.0 0.97 65.9 97.6

11 × 11 0.78 116 79.7 0.67 124 68.1 0.98 147 97.7

Table 2: Preceding and execution time analysis.

Map Size

GPT-4o Gemini-1.5 Claude-3.5

T T

5 × 5 84.6 2.93 81.5 107 3.20 104 85.9 2.81 83.1

7 × 7 129 4.94 124 125 3.67 121 130 3.18 127

11 × 11 169 9.47 160 157 5.56 151 184 5.84 178

a coverage rate above 65%. As shown in Table 1,

the model claude-3.5-sonnet exhibits the best perfor-

mance among the three models in terms of cover-

age rate and weighted path length. Changes in map

size do not signiﬁcantly affect the coverage rate and

weighted path for the model gemini-1.5-ﬂash. Con-

versely, the model gpt-4o achieves a higher coverage

rate with smaller map sizes, but this rate decreases

as the map size increases. As the map size grows,

the actual path length increases more rapidly than

the weighted path length, indicating that the planned

paths include repeated visits to the same cells based

on the random start position.

The differences in path length are attributed to

the coverage rate of the planned path and the mobile

agent’s hardware capabilities, such as sensors and ac-

tuators. Since the evaluation processes locally with

a short time cost (less than 300ms), we sum the in-

ference time and the evaluation time as T

. T and T

represent the total time spent and the driving part time

cost, respectively.

Model claude-3.5-sonnet performs best and ex-

hibits the fastest inference time in the experiment,

planning fully coverage waypoints in various envi-

ronments. Model gpt-4o shows stable performance

across different map sizes, demonstrating robustness

and reliability. However, it is noted that the model’s

performance declines slightly as the map size in-

creases, which could be attributed to the complex-

ity of managing larger spaces and more waypoints.

Model gemini-1.5-ﬂash, on the other hand, main-

tains consistent performance regardless of map size,

although it occasionally introduces extra line break

marks in its responses, which could be due to format-

ting issues within the LLM’s output generation pro-

cess.

Additionally, the path length differences highlight

the varying capabilities of the mobile agents’ hard-

ware, such as sensor accuracy and actuator precision,

which directly impact the execution of the planned

paths. The evaluation process, which includes both

inference and validation, ensures that the paths are not

only feasible but also optimized for efﬁciency.

Overall, the claude-3.5-sonnet model excels in

both performance and speed, making it ideal for sce-

narios requiring rapid and thorough coverage. The

gpt-4o model offers balanced performance with sta-

bility across various map sizes, making it a versatile

choice. The gemini-1.5-ﬂash model, despite minor

formatting issues, proves to be reliable with consis-

tent performance. These insights can guide the selec-

tion of appropriate LLM services for speciﬁc cover-

age path planning tasks in mobile robotics.

5 CONCLUSIONS

We propose a novel embodied framework for mobile

agents, incorporating weighted evaluation metrics for

the speciﬁc task of coverage path planning. A key fac-

tor of the framework is the use of zero-shot prompts to

simplify LLM inference during the initial phase. This

approach leverages the power of LLMs to generate ef-

fective waypoints without the need for extensive train-

ing data, thus streamlining the path-planning process.

During the navigation phase, we introduced a robust

safety mechanism for mobile agents to avoid obsta-

cles. This mechanism ensures that the mobile agents

can navigate safely and efﬁciently in dynamic envi-

ronments. Our experiments demonstrate that current

LLMs have the capability to function as an embod-

ied AI brain within mobile agents for speciﬁc tasks,

such as area coverage, when guided by appropriately

designed prompts.

The competition among LLM companies has sig-

niﬁcantly advanced the ﬁeld, freeing researchers from

the traditional labelling-training-validation loop in AI

research. This shift allows for more focus on inno-

SIMULTECH 2025 - 15th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

190

vative applications and real-world deployment of AI

technologies. Future research will focus on evaluat-

ing path-planning problems in more realistic scenar-

ios and simulation environments. This includes in-

tegrating more complex environmental variables and

constraints to further evaluate and enhance the robust-

ness of the proposed framework. Additionally, ex-

ploring the scalability of LLMs in diverse and larger-

scale applications will be crucial in advancing the

practical deployment of embodied AI systems in mo-

bile robotics.

REFERENCES

Ab Wahab, M. N., Nazir, A., Khalil, A., Ho, W. J., Ak-

bar, M. F., Noor, M. H. M., and Mohamed, A. S. A.

(2024). Improved genetic algorithm for mobile robot

path planning in static environments. Expert Systems

with Applications, 249:123762.

Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I.,

Aleman, F. L., Almeida, D., Altenschmidt, J., Altman,

S., Anadkat, S., et al. (2023). Gpt-4 technical report.

arXiv preprint arXiv:2303.08774.

Ali, Z. A. and Israr, A. (2024). Motion Planning for Dy-

namic Agents. BoD–Books on Demand.

Anderson, P., Chang, A., Chaplot, D. S., Dosovitskiy, A.,

Gupta, S., Koltun, V., Kosecka, J., Malik, J., Mot-

taghi, R., Savva, M., et al. (2018). On evalua-

tion of embodied navigation agents. arXiv preprint

arXiv:1807.06757.

Barraquand, J., Langlois, B., and Latombe, J.-C. (1992).

Numerical potential ﬁeld techniques for robot path

planning. IEEE transactions on systems, man, and

cybernetics, 22(2):224–241.

aunl, T. (2020). Robot adventures in Python and C.

Springer.

aunl, T. (2023). Mobile Robot Programming: Adventures

in Python and C. Springer International Publishing.

Cao, L. (2024). Ai robots and humanoid ai: Re-

view, perspectives and directions. arXiv preprint

arXiv:2405.15775.

Castillo, O., Trujillo, L., and Melin, P. (2007). Multiple ob-

jective genetic algorithms for path-planning optimiza-

tion in autonomous mobile robots. Soft Computing,

11:269–279.

Chen, Y., Han, Y., and Li, X. (2025). Fastnav: Fine-tuned

adaptive small-language- models trained for multi-

point robot navigation. IEEE Robotics and Automa-

tion Letters, 10(1):390–397.

Chrisley, R. (2003). Embodied artiﬁcial intelligence. Arti-

ﬁcial intelligence, 149(1):131–150.

Dewang, H. S., Mohanty, P. K., and Kundu, S. (2018). A

robust path planning for mobile robot using smart par-

ticle swarm optimization. Procedia computer science,

133:290–297.

Di Franco, C. and Buttazzo, G. (2016). Coverage path plan-

ning for uavs photogrammetry with energy and reso-

lution constraints. Journal of Intelligent & Robotic

Systems, 83:445–462.

Dorbala, V. S., Chowdhury, S., and Manocha, D. (2024).

Can llms generate human-like wayﬁnding instruc-

tions? towards platform-agnostic embodied instruc-

tion synthesis. arXiv preprint arXiv:2403.11487.

Driess, D., Xia, F., Sajjadi, M. S. M., Lynch, C., Chowd-

hery, A., Ichter, B., Wahid, A., Tompson, J., Vuong,

Q., Yu, T., Huang, W., Chebotar, Y., Sermanet, P.,

Duckworth, D., Levine, S., Vanhoucke, V., Haus-

man, K., Toussaint, M., Greff, K., Zeng, A., Mor-

datch, I., and Florence, P. (2023). Palm-e: An em-

bodied multimodal language model. In arXiv preprint

arXiv:2303.03378.

Ferguson, D. and Stentz, A. (2005). The ﬁeld d* algo-

rithm for improved path planning and replanning in

uniform and non-uniform cost environments. Robotics

Institute, Carnegie Mellon University, Pittsburgh, PA,

Tech. Rep. CMU-RI-TR-05-19.

Galceran, E. and Carreras, M. (2012). Efﬁcient seabed

coverage path planning for asvs and auvs. In 2012

IEEE/RSJ International Conference on Intelligent

Robots and Systems, pages 88–93. IEEE.

Gu, S. (2023). Llms as potential brainstorming part-

ners for math and science problems. arXiv preprint

arXiv:2310.10677.

Hadi, M. U., Qureshi, R., Shah, A., Irfan, M., Zafar, A.,

Shaikh, M. B., Akhtar, N., Wu, J., Mirjalili, S., et al.

(2023). A survey on large language models: Appli-

cations, challenges, limitations, and practical usage.

Authorea Preprints.

Hazem, S., Mostafa, M., Mohamed, E., Hesham, M., Mo-

hamed, A., Lotfy, E., Mahmoud, A., and Yacoub, M.

(2021). Design and path planning of autonomous so-

lar lawn mower. In International Design Engineering

Technical Conferences and Computers and Informa-

tion in Engineering Conference, volume 85369, page

V001T01A016. American Society of Mechanical En-

gineers.

Hewawasam, H., Ibrahim, M. Y., and Appuhamillage, G. K.

(2022). Past, present and future of path-planning al-

gorithms for mobile robot navigation in dynamic envi-

ronments. IEEE Open Journal of the Industrial Elec-

tronics Society, 3:353–365.

Hoffman, K. L., Padberg, M., Rinaldi, G., et al. (2013).

Traveling salesman problem. Encyclopedia of op-

erations research and management science, 1:1573–

1578.

Kuo, Y.-L., Katz, B., and Barbu, A. (2020). Deep compo-

sitional robotic planners that follow natural language

commands.

Latif, E. (2024). 3p-llm: Probabilistic path planning using

large language model for autonomous robot naviga-

tion.

Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter,

B., Florence, P., and Zeng, A. (2023). Code as poli-

cies: Language model programs for embodied control.

Ma, Y. J., Liang, W., Wang, G., Huang, D.-A., Bastani, O.,

Jayaraman, D., Zhu, Y., Fan, L., and Anandkumar, A.

Embodied AI in Mobile Robot Simulation with EyeSim: Coverage Path Planning with Large Language Models

191

(2024). Eureka: Human-level reward design via cod-

ing large language models.

Palacios-Gas

os, J. M., Sag

ues Blazquiz, C., and Monti-

jano Mu

noz, E. (2018). Multi-Robot Persistent Cover-

age in Complex Environments. PhD thesis, PhD thesis,

Universidad de Zaragoza.

Panov, A. I., Yakovlev, K. S., and Suvorov, R. (2018).

Grid path planning with deep reinforcement learn-

ing: Preliminary results. Procedia computer science,

123:347–353.

Petitjean, S. (2002). A survey of methods for recovering

quadrics in triangle meshes. ACM Computing Surveys

(CSUR), 34(2):211–262.

Schumann, R., Zhu, W., Feng, W., Fu, T.-J., Riezler, S., and

Wang, W. Y. (2024). Velma: Verbalization embodi-

ment of llm agents for vision and language navigation

in street view. Proceedings of the AAAI Conference

on Artiﬁcial Intelligence, 38(17):18924–18933.

Sharma, P., Sundaralingam, B., Blukis, V., Paxton, C.,

Hermans, T., Torralba, A., Andreas, J., and Fox, D.

(2022). Correcting robot plans with natural language

feedback.

Singh, I., Blukis, V., Mousavian, A., Goyal, A., Xu, D.,

Tremblay, J., Fox, D., Thomason, J., and Garg, A.

(2022). Progprompt: Generating situated robot task

plans using large language models.

Smith, D. R. (1985). The design of divide and conquer algo-

rithms. Science of Computer Programming, 5:37–58.

Su, J., Jiang, C., Jin, X., Qiao, Y., Xiao, T., Ma, H., Wei,

R., Jing, Z., Xu, J., and Lin, J. (2024). Large language

models for forecasting and anomaly detection: A sys-

tematic literature review.

Torres, M., Pelta, D. A., Verdegay, J. L., and Torres, J. C.

(2016). Coverage path planning with unmanned aerial

vehicles for 3d terrain reconstruction. Expert Systems

with Applications, 55:441–451.

Warren, C. W. (1993). Fast path planning using modiﬁed a*

method. In [1993] Proceedings IEEE International

Conference on Robotics and Automation, pages 662–

667. IEEE.

Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan,

K., and Cao, Y. (2022). React: Synergizing reason-

ing and acting in language models. arXiv preprint

arXiv:2210.03629.

Zhang, R., Jiang, D., Zhang, Y., Lin, H., Guo, Z., Qiu, P.,

Zhou, A., Lu, P., Chang, K.-W., Gao, P., et al. (2024).

Mathverse: Does your multi-modal llm truly see the

diagrams in visual math problems? arXiv preprint

arXiv:2403.14624.

Zhao, M., Anderson, P., Jain, V., Wang, S., Ku, A.,

Baldridge, J., and Ie, E. (2021). On the evaluation

of vision-and-language navigation instructions.

SIMULTECH 2025 - 15th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

192