Dynamic Two-Dimensional Maze Generation and Decryption with

Weights and Traps Based on DFS and Dijkstra Algorithm

Linjia Guo

Institute of Waterford, Nanjing University of Information Science and Technology, Ningliu Road, Nanjing, China

Keywords: Maze Generation, Dynamic Pathfinding, Depth-First Search, Dijkstra Algorithm, Robotics Navigation.

Abstract: Dynamic weighted maze generation and solving represent a critical challenge for intelligent navigation

systems in complex and unpredictable environments. In this study, the paper proposes a modular and

extensible platform that integrates randomized depth-first search (DFS) for unique maze construction and

supports real-time hazard injection, such as traps and mutable barriers. For pathfinding, both breadth-first

search (BFS) and Dijkstra’s algorithm are implemented to handle unweighted and weighted scenarios,

respectively, thereby enabling adaptive and cost-sensitive navigation. The solver is instantly re-invoked

whenever the maze environment changes, ensuring up-to-date risk avoidance and route optimization.

Extensive experiments in two-dimensional and three-dimensional mazes demonstrate that the proposed

platform achieves highly reliable, robust, and efficient navigation, even under frequent dynamic updates. The

system features real-time visualization, making it an ideal tool for research, algorithm benchmarking, and

education in areas such as robotics and intelligent path planning. This work provides a solid foundation for

further investigation into resilient autonomous navigation under real-world uncertainties.

1 INTRODUCTION

Maze generation and solving are fundamental topics

in artificial intelligence and robotics, widely

employed as abstract models for path planning and

navigation in complex, uncertain environments. In

practical applications, navigation is rarely static—

autonomous robots and emergency rescue teams

regularly confront evolving obstacles and hazardous

zones, demanding rapid response and adaptive

decision-making from intelligent systems.

Consequently, dynamically generated, weighted

mazes that incorporate traps and variable barriers

provide a far more realistic testbed for validating and

advancing intelligent pathfinding algorithms.

Recent advances underscore both the challenges

and significance of this problem. A quantum-walk-

based framework has excelled at finding shortest

paths in static, unweighted mazes, offering

compelling theoretical foundations (Matsuoka, Ohno,

& Segawa, 2025).Nevertheless, such approaches do

not address the intricacies of dynamically changing

environments or the effects of cell-specific risks.

Multi-agent maze exploration has also been enhanced

using improved flood-fill algorithms, emphasizing

coordination among agents in unknown environments

(Tjiharjadi, Razali, Sulaiman, 2022; Fernando, 2022).

However, these strategies omit dynamically shifting

hazards and weighted traversal costs. Hybrid

algorithms have been proposed for minimizing

expected escape time, balancing path length and risk,

though they do not consider mazes with dynamically

updated traps or barriers(Yendri, Soelaiman, &

Purwananto, 2023). Solutions that integrate ant

colony algorithms with Dijkstra’s algorithm for

search and rescue illustrate the potential for

combining exploration and hazard-avoidant

navigation(Husain, Zaabi, Hildmann, et al, 2022).

Still, such approaches typically treat navigation and

response to danger as separate phases rather than as

continuous, adaptive strategies.

Despite these developments, there remains a

notable gap, as existing research seldom integrates

dynamic, weighted hazards with pathfinding methods

capable of instant adaptation to environmental

changes. This is particularly pressing in domains such

as smart manufacturing and disaster relief, where

resilient navigation systems must simultaneously

ensure safety and efficiency under unpredictable

conditions. Addressing how to combine robust risk

avoidance with optimal path efficiency in the face of

Guo, L.

Dynamic Two-Dimensional Maze Generation and Decryption with Weights and Traps Based on DFS and Dijkstra Algorithm.

DOI: 10.5220/0014325700004718

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2025), pages 227-231

ISBN: 978-989-758-792-4

227

ongoing changes is key to advancing reliable

autonomous systems.

This project presents an extensible platform for

dynamic weighted maze generation and solving,

supporting both 2D and 3D scenarios. The system

employs depth-first search (DFS) to generate unique

maze solutions while dynamically injecting random

traps and variable barriers. For maze solving, the

framework integrates both breadth-first search (BFS)

for standard cases and a weight-sensitive Dijkstra

algorithm for efficient, adaptive path planning as

conditions change in real time. Integrated

visualization modules provide intuitive displays of

solution paths and risk distributions, making the

platform well-suited as a simulation and teaching tool

for intelligent robotics and rescue operations.

By uniting dynamic hazard modeling with real-

time, weight-adaptive pathfinding, this work not only

advances the theoretical foundation of intelligent

navigation but also delivers practical value for

research and real-world deployment in challenging

environments.

2 METHODOLOGY

To create a realistic and challenging pathfinding

environment, the paper model the maze as a grid-

based structure in either two or three dimensions.

Each grid cell is assigned a type: wall (impassable),

space (traversable), or trap (traversable but with a

significantly higher traversal cost) (Shi, & Ma. 2022;

Chen. 2020; Tang, Jiang, Gao, et al, 2024). The

internal data organization is a matrix of Cell objects

for 2D, or nested lists in 3D, where each Cell keeps

track of wall status, trap status, and its traversal

weight or cost. Upon initialization, all cells are made

walls and default weights are assigned, ready for

maze carving and hazard insertion.

Maze generation leverages the randomized DFS

algorithm, which is selected for its efficiency and its

guarantee to produce “perfect” mazes—connectivity

without cycles and a unique solution between any two

points (Shi, He, & Ma. 2022). The process begins by

selecting a random starting cell; from there, the

algorithm recursively explores all unvisited

neighbors in random order, removing walls between

the current cell and its neighbor to create legal

passages. When it reaches a dead-end, it backtracks

until another unexplored branch is found. In two

dimensions, neighbors are checked in the four

cardinal directions, while in three dimensions, up and

down are added for extra spatial complexity.

After maze construction, traps are randomly

placed to create high-cost traversal zones, mimicking

hazardous locations. Dynamic behavior is further

simulated by allowing traps or walls to emerge or

disappear during runtime, representing the

unpredictable changes encountered by robots and

agents in real environments (Husain, Zaabi,

Hildmann, et al, 2022; Tjiharjadi, Razali, Sulaiman,

et al, 2022).

For finding routes through the maze, the paper

implements both BFS and Dijkstra’s algorithm to

accommodate different scenario complexities. BFS is

primarily used for unweighted mazes where each step

has uniform cost, guaranteeing the derivation of the

minimal-step route from start to goal (Chen, 2020).

The algorithm enqueues the starting cell, explores all

reachable neighbors in layers.

When cell weights vary—due to traps or other

hazards—Dijkstra’s algorithm is employed as it seeks

the path with the minimum total cost, regardless of

spatial distance (Yendri, Soelaiman, & Purwananto,

2023; Husain, Zaabi, Hildmann, et al, 2022). This

algorithm maintains a priority queue, always

expanding the node with the lowest accumulated cost

so far and updating the cost of reaching each neighbor

when a less expensive route is discovered. Upon

reaching the goal, the total cost and the path can be

reconstructed efficiently.

In both algorithms, each time the maze changes

during simulation (e.g., when a new trap appears or a

wall is erected), the corresponding solver is invoked

anew so that the navigation path reflects the latest

maze structure and hazards (Montenegro, Robinson,

Fu, et al, 2025; Valenzuela, Schaa, Barriga, et al,

2022). Visualization is handled using matplotlib with

black lines indicating walls, red hues for traps, and

blue lines highlighting the final chosen path. During

experimentation, the paper test on grid sizes such as

20×15 for 2D and up to 5×5×5 for 3D, with hundreds

of random dynamic events per simulation. Key

performance metrics—including path length, total

cost, traps avoided, and the speed of adaptation to

changes—are logged throughout each run.

3 RESULTS

The effectiveness and adaptability of our dynamic

weighted maze generation and solving system are

vividly illustrated in Figure 1, which showcases a

typical experimental output from our platform. In this

2D case, the maze is initially constructed using the

randomized DFS algorithm, ensuring a uniquely

solvable structure without loops (Shi, He, & Ma,

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

228

2022). The maze’s intricate walls, drawn in black,

create a challenging yet fully traversable environment

from the designated start in the upper left to the exit

at the bottom right.

Figure 1: Generated maze display diagram. (Picture credit:

Original)

Within the maze, several cells are highlighted in

translucent red, representing traps or hazard zones.

These special cells were introduced after the basic

maze construction and assigned a higher traversal

cost, emulating real-world dangers or impediments an

autonomous agent might encounter in unpredictable

terrain. The location of each trap is determined

randomly at runtime. Furthermore, during each

simulation step, trap status or even the presence of

certain walls may be adjusted dynamically to

replicate the evolving hazards typical in actual

robotics navigation, such as sudden blockages or new

risk zones (Husain, Zaabi, Hildmann, et al, 2022).

Pathfinding within this dynamic environment is

performed using Dijkstra’s algorithm, which is

particularly well-suited for weighted mazes with

varying traversal costs. In the displayed example, the

solution path is highlighted with a bold blue line.

Unlike simple shortest-path algorithms such as BFS,

Dijkstra’s algorithm seeks to minimize total traversal

cost rather than the sheer number of steps. As such,

the calculated route purposely circumvents most red

trap areas, sometimes accepting a longer geometric

pathway to ensure the agent avoids incurring high risk

or cost (Yendri, Soelaiman, & Purwananto, 2023;

Husain, Zaabi, Hildmann, et al, 2022). This behavior

is clearly observed in Figure 1: the blue route

consistently detours around trap zones, adhering

closely to safe corridors even if it means not taking

the most direct route to the goal.

The dynamic aspect of the system is further

demonstrated throughout the simulation cycles.

When new traps or barriers are triggered during

navigation, the algorithm promptly recalculates and

adapts the current route. This allows the agent to

respond instantly to changing hazards, a capability

essential for real-time robotics navigation and search-

and-rescue applications (Montenegro, Robinson, Fu,

et al,2025; Valenzuela, Schaa, Barriga, et al, 2022).

Repeated experiments confirm the platform’s

stability and responsiveness, with path recalculations

occurring almost instantly as the maze environment

mutates.

Quantitatively, the system yields several clear

advantages. For mazes without traps (where all

weights are uniform), Dijkstra’s and BFS perform

equivalently, both returning the shortest possible step

path (Yendri, Soelaiman, & Purwananto, 2023; Chen,

2020). However, when hazards are introduced,

Dijkstra’s approach consistently reduces the number

of risky (trap-laden) cells traversed, often at the cost

of additional path length. This tradeoff mirrors the

demands of realistic environments, where safe

navigation must be prioritized above sheer efficiency.

Our experiments further show that, regardless of the

frequency and intensity of dynamic updates, the

system reliably avoids new hazards and maintains

high-quality solution paths, supporting ongoing

adaptation as new obstacles or risks appear.

Lastly, the integration of real-time visualization

not only makes the system accessible for educational

and research purposes, but also facilitates

troubleshooting and performance assessment

(Montenegro, Robinson, Fu, et al, 2025). The

immediate graphical feedback allows users to observe

precisely how the agent negotiates traps and barriers,

and to verify that the implemented algorithms behave

as intended under a wide variety of challenging and

changing scenarios (Valenzuela, Schaa, Barriga, et al,

2022).

In conclusion, the results captured in Figure 1

exemplify the strengths of the dynamic weighted

maze platform: robust organic maze generation,

intelligent risk-aware navigation using Dijkstra’s

algorithm, and thorough visual feedback—all of

which support robust benchmarking and exploration

of intelligent path planning under realistic, uncertain

conditions.

4 DISCUSSION

The results underscore the necessity and advantages

of employing dynamic, weighted maze models for

testing and advancing intelligent navigation. The

DFS-based generator provides an unending supply of

challenging, uniquely solvable environments that

maintain their value across both 2D and 3D cases,

Dynamic Two-Dimensional Maze Generation and Decryption with Weights and Traps Based on DFS and Dijkstra Algorithm

229

aligning well with trends in game design, simulation,

and evaluation methods favored in recent works

(Montenegro, Robinson, Fu, et al, 2025; Valenzuela,

Schaa, Barriga, & Patow, 2022). By supporting not

only static but also endlessly changing configurations,

the system closely matches the shifting hazards found

in real robotics, public safety, and emergency

applications.

However, limitations remain. The current model

assigns traps randomly, which does not fully capture

the spatial correlations or sensor-informed risks

present in true operational contexts. The visualization

module, though clear, could be further developed for

3D interactive inspection and overlay of additional

data, such as energy use or threat probability fields.

Most notably, while Dijkstra’s solver is effective and

robust, it recalculates from scratch after each update.

Future research should consider incremental or

heuristic approaches that reuse prior computations to

further reduce latency, as well as hybrid multi-agent

or probabilistic path planning strategies inspired by

recent industry advances (Tjiharjadi, Razali,

Sulaiman, 2022; Fernando, 2022; Yendri, Soelaiman,

& Purwananto, 2023; Husain, Zaabi, Hildmann, et al,

2022).

Comparing our results to the literature, the

presented system extends the state of the art by

facilitating continuous, online adaptation in the face

of dynamic hazards, a clear step beyond most prior

work emphasizing static or phase-based approaches

(Matsuoka, Ohno, & Segawa, 2025; Matsuoka, Yuki,

Lavička, & Segawa, 2021; Husain, Zaabi, Hildmann,

et al, 2022). The easily extensible simulation and

visualization environment also offers a strong

foundation for deployment in education, algorithm

research, or hardware-in-the-loop validation.

Opportunities for further extension include real-

world data integration, richer user interaction, and

cross-validation with quantum-inspired or multi-

agent approaches documented in current academic

research (Montenegro, Robinson, Fu, et al, 2025;

Valenzuela, Schaa, Barriga, & Patow, 2022). This

work thus lays a solid groundwork for safer and

smarter autonomous navigation in increasingly

complex and dynamic real-world environments.

5 CONCLUSION

This paper presents a comprehensive solution for

dynamic weighted maze generation and adaptive

pathfinding, addressing a core need in intelligent

robotics and navigation research. By combining

randomized DFS maze generation with real-time risk

modeling and integrating both BFS and Dijkstra’s

algorithms, the proposed platform robustly adapts to

continuous environmental changes. Experimental

results confirm that the system maintains high

efficiency and safety when navigating evolving

mazes, consistently avoiding newly introduced traps

and overcoming unexpected obstacles. The

platform’s real-time visualization and modular design

not only facilitate academic exploration and

educational use, but also offer broad applicability for

future expansion—such as the integration of sensor-

driven risk assessment, incremental search strategies,

and multi-agent coordination. Overall, this research

provides valuable practical and theoretical

contributions towards the development of reliable,

resilient, and intelligent autonomous navigation in

complex and dynamic environments.

REFERENCES

Chen, J. (2020). Implementation of maze generation and

pathfinding program based on Scratch 3.0. Digital

Technology and Applications, 38(9), 96–98.

Husain, Z., Al Zaabi, A., Hildmann, H., Saffre, F., Ruta, D.,

& Isakovic, A. F. (2022). Search and rescue in a maze-

like environment with ant and Dijkstra algorithms.

Drones, 6(10), 273.

Matsuoka, L., Ohno, H., & Segawa, E. (2025). A

mathematical framework for maze solving using

quantum walks. Quantum Information Processing, 24,

93.

Montenegro, G. A. D. C., Robinson, T. L., Fu, J. F.,

Properzi, M. J., Penney, D., Davis, R., Sperling, R. A.,

Johnson, K. A., Amariglio, R. E., & Rentz, D. M.

(2025). The digital maze test offers a unique method for

detecting early cognitive changes in preclinical

Alzheimer’s disease. Alzheimer’s & Dementia, 20

(Suppl8), e095237.

Matsuoka, L., Yuki, K., Lavička, H., & Segawa, E. (2021).

Maze solving by a quantum walk with sinks and self-

loops: Numerical analysis. Symmetry, 13(12), 2263.

Shi, B., He, Y., & Ma, S. (2022). Research and design of

improved disjoint-set-based maze map generation

algorithm. Journal of Changchun Normal University,

41(4), 51–55.

Tang, D., Jiang, E., Gao, X., & others. (2024). Design and

modeling of modular three-dimensional maze

amusement facilities. Journal of Hubei University of

Technology, 39(6), 106–112.

Tjiharjadi, S., Razali, S., Sulaiman, H., & Fernando, G. V.

(2022). Design of multi-agent pathfinding robot using

improved flood fill algorithm in maze exploration.

International Journal of Mechanical Engineering and

Robotics Research, 11(8), 631–638.

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

230

Valenzuela, E., Schaa, H., Barriga, N., & Patow, G. (2022).

Using search algorithm statistics for predicting puzzle

difficulty.

Yendri, S., Soelaiman, R., & Purwananto, Y. (2023).

Hybrid algorithm to find minimum expected escape

time from a maze. Engineering Letters, 31(1), 346–357.

Dynamic Two-Dimensional Maze Generation and Decryption with Weights and Traps Based on DFS and Dijkstra Algorithm

231