CoMaDNav : Collaborative Dynamic Navigation

and Mapping in Autonomous Warehouse

Environments: A Multi-Agent Approach Using
DWA and Global Replanning
Swayam Rajat Mohanty Ankesh Ansh
Electrical Engineering Electrical Engineering
IIT BHU IIT BHU
22085104 22085016

Abstract—This paper introduces a simulation framework de-

signed for multi-robot navigation in dynamic warehouse settings.
It cleverly combines a Dynamic Window Approach (DWA)
with global re-planning using A* search. Our algorithm takes
advantage of a shared global occupancy map, which is built
from the combined LIDAR data of several robots. This allows
for continuous updates to the representation of both static
obstacles, like shelves and walls, and dynamic ones, such as
moving objects. Each robot employs local planning through
DWA for quick navigation. However, if a robot notices it’s not
making enough progress toward its goal, it switches to an A*-
based global planner to find a new route. We assess the method
using various performance metrics, including time-to-goal, total
distance traveled, path efficiency, clearance, and computational
load, across extensive simulation runs in a large-scale grid
environment.

I. I NTRODUCTION
The rapid growth of e-commerce and automated logistics
has driven the need for efficient and reliable multi-robot Fig. 1: Multiple bots path planing and tracking along with
navigation systems in dynamic warehouse environments mapping of the dynamically changing environment
[1]. Modern warehouses are quite intricate, filled with a blend
of static obstacles like shelving units and walls, alongside
dynamic ones such as other robots and human workers. moving objects. Each robot employs local planning through
Navigating these environments safely and efficiently, without DWA for quick navigation. However, if a robot notices it’s
collisions, is no small feat. It demands a thoughtful approach not making enough progress toward its goal, it switches to an
that combines global path planning, local obstacle avoidance, A*-based global planner to find a new route. We assess the
and the ability to adapt in real-time. method using various performance metrics, including time-
This paper introduces a novel simulation framework to-goal, total distance traveled, path efficiency, clearance,
for multi-robot navigation that addresses these challenges and computational load, across extensive simulation runs in
by integrating the Dynamic Window Approach (DWA) a large-scale grid environment. In dynamic and cluttered
with global re-planning based on A* search. This paper environments such as large-scale warehouses, autonomous
introduces a simulation framework designed for multi-robot multi-robot systems are expected to navigate efficiently
navigation in dynamic warehouse settings. It cleverly com- while avoiding collisions with both static obstacles (e.g.,
bines a Dynamic Window Approach (DWA) with global re- shelves and walls) and dynamic obstacles (e.g., moving
planning using A* search. Our algorithm takes advantage objects or people). Traditional multi-robot systems often rely
of a shared global occupancy map, which is built from on inter-robot communication to coordinate movements and
the combined LIDAR data of several robots. This allows share sensor information [2]. Relying on this approach does
for continuous updates to the representation of both static come with its fair share of challenges. For one, communication
obstacles, like shelves and walls, and dynamic ones, such as can be spotty in areas plagued by radio frequency interference
or physical barriers. Plus, it often requires a hefty amount of efficiency, and suitability for complex maps.” In a similar vein,
computational and network resources, which can be a burden. the Multi-Robot Rapidly exploring Random Tree (MRRRT)
There’s also the risk of it becoming a single point of failure. In has found its place in multi-robot scenarios. However, Alotaibi
situations where safety is critical—like search and rescue mis- and Al-Rawi point out that while MRRRT is great at exploring
sions, warehouse logistics, and human-robot interactions—it’s search spaces and optimizing solutions, it might not be as ef-
absolutely vital to ensure that navigation remains robust and fective as other methods when it comes to resolving collisions.
fault-tolerant, all without relying on constant communication. [7].
[3], [4].
At the core of our framework is a two-tiered planning B. Graph-Based and Network Approaches
strategy. Each robot employs DWA for local planning, allow- Recent advancements have brought sophisticated network
ing for reactive navigation and collision avoidance in real time. designs into the realm of path planning. The Multi-Agent
In parallel, the system continuously updates a shared global Routing in Variable Environments with Learning (MARVEL)
occupancy map derived from the collective LIDAR data of framework leverages Graph Attention Networks (GATs) and
all robots. This global map is crucial as it provides up-to- deep reinforcement learning to effectively navigate through
date, comprehensive information on both static and dynamic unpredictable topological settings. [8]. This method takes
obstacles, thereby informing and enhancing both the local and the navigation challenge and reimagines it as a Partially
global planning processes. When a robot detects insufficient Observable Markov Decision Process (POMDP). It also uses
progress toward its goal—due, for instance, to congestion or self-attention mechanisms to hone in on the most important
blocked paths—it autonomously triggers a global re-planning features of the graph.
process using the A* algorithm to compute an alternative
route. This combination of reactive local navigation, real-time C. Hybrid Local-Global Planning Approaches
global map updates, and adaptive global planning ensures Hybrid methods that blend global path planning with local
robust performance in highly dynamic scenarios while min- obstacle avoidance are becoming increasingly popular for real-
imizing the reliance on explicit inter-robot communication. world applications. One interesting approach is the integration
The decision to forgo explicit intercommunication in our of the A* algorithm with the Dynamic Window Approach (IA-
multi-robot system is motivated by several practical con- DWA), which aims to ”quickly search for globally optimal
siderations. First, in many real-world indoor environments, collision-free paths and avoid unknown obstacles in time”
reliable communication is difficult to guarantee due to signal [9]. On a similar note, the Multi-module Enhanced DWA
interference and physical obstructions [4]. Second, eliminating (MEDWA) tackles the challenges faced by standard DWA in
the need for continuous data exchange simplifies the system ar- crowded environments by using vector field techniques and
chitecture and reduces the computational overhead associated dynamically adjusting the weights of the evaluation func-
with processing and integrating external messages, thereby tions. [10]. Another notable hybrid approach is the ACO-
increasing overall robustness and scalability [5]. Finally, a DWA method, which combines Ant Colony Optimization for
non-communicative design enhances fault tolerance: even if global planning with the Dynamic Window Approach for local
a robot’s sensors or computational unit temporarily fail, the navigation [11].
remaining robots can continue operating autonomously with-
out waiting for central coordination or relying on potentially D. Map Sharing and Environmental Awareness
compromised communication links. Effective environmental representation and sharing remain
Through extensive simulation experiments, we evaluate our crucial challenges in multi-robot systems. Techniques for
algorithm using key performance metrics—including time- merging occupancy grid maps from multiple robots have
to-goal, total distance traveled, path efficiency, minimum enabled teams to build comprehensive environmental models
clearance, and computational load—to demonstrate that the without prior knowledge of relative positions [12]. This ap-
combined local-global planning approach not only ensures safe proach uses similarity metrics and stochastic search algorithms
and efficient navigation in complex warehouse scenarios but to identify overlapping regions and join local maps into a
also leverages the inherent benefits of a communication-free cohesive global representation. In addition, maintaining line-
multi-robot system. of-sight connectivity in environments with unknown obstacles
has been addressed using point cloud data processing and
II. L ITERATURE REVIEW
potential functions based on graph connectivity metrics [13].
A. Sampling-Based Approaches
Sampling-based methods have really taken off in the world E. Coordination Strategies
of multi-robot path planning. One standout in this area is Multi-robot coordination strategies have evolved from sim-
the Multi-Agent Rapidly-exploring Random Disjointed-Trees ple priority-based approaches to sophisticated distributed sys-
(MA-RRdT*) algorithm, which marks a significant step for- tems. As Parker details, coordination methods can be cate-
ward. This was highlighted by Zhang and colleagues. [6], gorized based on their relation to other robots (formations,
MA-RRdT* uses a method called multi-tree random sampling, flocking), the environment (search, coverage), or external
which helps it achieve “fast speed, high space exploration agents [14]. According to Bui’s survey, multi-robot motion
 Fig. 2: Detailed flowchart of the proposed algorithm.

planning taxonomies can be organized by algorithm character- • Their positions are discretized and combined into a grid
istics, approaches, and application scenarios [15]. Challenges representation (grid_dynamic).
remain in scaling to larger teams, handling highly stochastic • The static and dynamic grids are merged (using a cell-
environments, and integrating motion and sensing constraints. wise maximum) to form the overall occupancy grid
(grid_combined_all).
III. D ETAILED E XPLANATION OF THE A LGORITHM
C. Global Map Update
Our proposed algorithm is designed for multi-robot nav-
• Each robot uses its LIDAR sensor to obtain range data
igation in dynamic warehouse environments and combines
at multiple angles.
reactive local planning with adaptive global re-planning. The
• The sensor readings are used to update the global oc-
overall strategy is decentralized and does not rely on inter-
cupancy map, marking detected obstacles. This global
robot communication; instead, each robot uses its own sensor
map provides a richer, more robust representation of the
data to build a local representation of the environment while
environment, allowing both local and global planners to
benefiting from a shared global occupancy map that is con-
operate on up-to-date information.
tinuously updated from collective LIDAR data. The algorithm
comprises the following key components: D. Robot-Specific Processing
For every robot in the simulation, the following sequence
A. Environment and Map Initialization
is executed:
• Static Grid Setup: The environment is represented 1) Goal Check: The robot first checks if it has reached
as a two-dimensional grid. Static obstacles (e.g., its main goal (using a predefined threshold). If so, it
shelving units, walls) are encoded in a binary grid is marked as finished and the time-to-goal metric is
(grid_static), where occupied cells indicate the pres- recorded.
ence of obstacles. 2) Local Goal Determination: If the robot is not yet at its
• Global Occupancy Map: A global occupancy map is main goal, it determines its current planning target. This
initialized to record environmental changes. This map is either the main goal (in the absence of obstacles) or
is updated continuously based on sensor readings from a subgoal extracted from a previously computed global
all robots, thereby capturing both static and dynamic path (if the robot is in “global mode”).
obstacles. 3) Local Planning via DWA:
• The Dynamic Window Approach (DWA) is em-
B. Dynamic Obstacles Update
ployed to compute the control inputs (linear and
At every simulation iteration, the algorithm updates the state angular velocities) based on the robot’s current state
of dynamic obstacles: and the local map.
• Each dynamic obstacle has a position and velocity, which • The DWA module iterates over possible velocity
are updated using a simple motion model. pairs and simulates short-term trajectories.
 • For each simulated trajectory, a cost is computed Metric Layout 1 Layout 2 Layout 3
which includes terms for the distance to the current ComadNav
goal, heading deviation, velocity performance, and Time to goal (s) 72.20 49 67
Total distance traveled 70.11 47.27 65.51
clearance from obstacles. The trajectory with the Path efficiency (ratio) 1.15 1.10 1.07
lowest cost (if collision-free) is selected. Cumulative heading change (rad) 9.10 6.69 4.67
Simple DWA
4) Stuck Detection and Global Replanning: Time to goal (s) 75.40 68.6 67.2
Total distance traveled 69.52 66.5 65.7
• The algorithm monitors the robot’s progress toward Path efficiency (ratio) 1.14 1.56 1.10
its goal. If the change in distance is below a thresh- Cumulative heading change (rad) 7.01 10.97 4.68
old over several iterations (indicating that the robot ACO DWA
might be stuck due to congestion or a local minima), Time to goal (s) 78.00 52.10 78.6
a counter is incremented. Total distance traveled 74.03 49.54 75.17
Path efficiency (ratio) 1.22 1.16 1.24
• Once the counter exceeds a set limit, the robot Cumulative heading change (rad) 8.89 9.37 14.36
switches to “global mode.” In this mode, the al-
A* with PID
gorithm invokes a global planner based on the A* Time to goal (s) 61.30 43.35 -
search algorithm over the current informed grid. Total distance traveled 61.05 43.05 -
The planner computes a collision-free path from the Path efficiency (ratio) 1.02 1.01 -
Cumulative heading change (rad) 23.75 36.27 -
robot’s current cell to the goal cell.
• The computed global path is segmented into sub- TABLE I: The table compares the performance of CoMadNav
goals. The robot then follows these subgoals until to standalone DWA, ACO DWA and A* with PID. THrought
it either reaches the main goal or the global mode the comparison of these metrics we can clearly see that the
times out. proposed algorithm outperforms existing algorithms.
5) State Update: The robot’s state is updated using simple
kinematic equations:
IV. P ERFORMANCE COMPARISON
To rigorously assess the effectiveness of our proposed path
xnew = xold + v cos(θ) ∆t, planning algorithm, we first establish baseline performance by
comparing it with two widely-adopted methods: the standalone
ynew = yold + v sin(θ) ∆t, Dynamic Window Approach (DWA) and the standalone A*
search algorithm. Standalone DWA is well-regarded for its
θnew = θold + ω ∆t.
efficiency in reactive local obstacle avoidance and real-time
responsiveness. In contrast, the A* algorithm is traditionally
valued for its ability to compute optimal paths on discretized
where v and ω are the selected linear and angular grids in static environments. Along side this we also compare
velocities, respectively, and ∆t is the time step. our algorithm to a widely adopted algorithm name Ant colony
6) Metrics and Visualization: Throughout each iteration, optimization(ACO) with DWA [16]. The method, though sim-
key metrics such as total distance traveled, heading ilar to ours, has some fundamental differences, such as we use
changes, minimum clearance, and computation time per A* for global re-planning when DWA gets stuck; the ACO-
iteration are recorded. An active visualization module DWA method uses ACO to generate the path and then DWA
updates the display, showing the occupancy maps, robot to track that path.
trajectories, current positions, goals, and global paths (if Our evaluation considers key performance metrics including
applicable). time-to-goal, total distance traveled, and path efficiency.
• Time to goal (s): The total time in seconds required for
the robot to reach its destination.
• Total distance traveled: The overall distance covered by
E. Termination and Post-Processing the robot along its path.
• Path efficiency (ratio): A ratio of the actual path length

• The simulation loop continues until all robots have to the optimal (shortest) path length; values closer to 1
reached their goals or the maximum number of iterations indicate more efficient paths.
is reached. • Cumulative heading change (rad): The sum of all

• Upon termination, the algorithm post-processes the angular changes (in radians) during the robot’s move-
recorded metrics, computing path efficiency (the ratio ment, which provides insight into the smoothness of the
of actual distance traveled to the optimal straight-line trajectory.
distance), average heading change, average velocity, and In Layout1 (a simple warehouse, see Fig 3 and TableI),
overall computation time. all four methods reach the goal reliably, but A* with PID
Fig. 3: Performance comparison of the four algorithms (CoMadNav, DWA, ACO DWA, and A with PID) across three warehouse
layouts with increasing complexity.* Layout 1 (a, b, c, d) represents a simple warehouse environment. Layout 2 (e, f, g, h)
introduces additional complexity, where it is evident that the basic DWA algorithm becomes trapped in a local minimum,
resulting in a non-optimal path. Layout 3 (i, j, k) further increases the complexity by introducing dynamic obstacles. In this
scenario, the A* with PID algorithm fails to perform effectively due to its inability to detect and respond to dynamic changes
in the environment.

is fastest. Because it follows a precomputed path with- smooth motion primitives and invokes A* only when necessary
out processing LIDAR data in real time, it incurs virtu- to escape stagnation, yielding both a fluid path and shorter
ally no online computation during tracking. The remain- time-to-goal.
ing three—ACODWA, Simple DWA, and our proposed Co- Layout 3 introduces moving obstacles (blue in Fig. 4) within
madNav—perform comparably, with ComadNav’s travel time the complex warehouse. Here, A* with PID fails outright,
falling midway between the two DWA-based baselines. as it cannot detect or react to dynamic obstacles. Both ACO
When we increase corridor complexity in Layout 2, per- DWA and Simple DWA struggle under these conditions: ACO
formance diverges markedly. Simple DWA (Fig. 3f) becomes DWA’s uneven waypoint sequence degrades efficiency, while
trapped in a local minimum and loops repeatedly, owing to Simple DWA again succumbs to local minima. In contrast,
its reliance on a finely tuned cost function to handle sharp ComadNav integrates real-time obstacle sensing and dynamic
direction changes. ACODWA avoids the deadlock but gen- replanning, successfully avoiding moving obstacles and main-
erates a jagged trajectory—its ant-colony optimizer produces taining a near-optimal trajectory.
suboptimal waypoints, resulting in large heading adjustments Overall, these experiments demonstrate that ComadNav
and slower progress. By contrast, ComadNav leverages DWA’s combines the computational efficiency of DWA with the
 Fig. 4: Detection and avoidance of dynamic obstacles using ComadNav

global foresight of A*, making it particularly well suited replanning—provides a balanced solution to the exploration-
for dynamic warehouse environments. exploitation dilemma in path planning. While several existing
approaches, such as IA-DWA [9] and ACO-DWA [11] also
A. Advantages of the Proposed Hybrid DWA-A* Approach combine local and global planning, the proposed algorithm’s
with Shared Occupancy Mapping trigger mechanism for global replanning based on progress
The proposed algorithm presents several potential advan- assessment is particularly well-suited for warehouse environ-
tages over existing approaches for warehouse environments: ments. When a robot detects insufficient progress toward its
1) Integrated Environmental Awareness: The shared goal, the A*-based global planner computes an alternative
global occupancy map, constructed from collective LIDAR route, effectively addressing situations where robots might
data, represents a significant advancement over approaches that become trapped in local minima or deadlocks due to dynamic
rely solely on individual robot perceptions. By continuously obstacles or congestion.
updating both static and dynamic obstacles, the system main-
V. D ISCUSSION
tains a comprehensive environmental model that benefits all
robots. This approach aligns with the work of Birk and Carpin This algorithm seamlessly integrates reactive local planning
on map merging [12] but extends it to real-time dynamic with adaptive global re-planning while continuously updating
updates, which is critical in warehouse environments where a shared global occupancy map. The decision to forgo ex-
inventory, equipment, and personnel are constantly moving. plicit inter-robot communication reduces system complexity
2) Adaptive Planning Hierarchy: The two-tiered planning and enhances robustness, critical advantages in dynamic and
approach—combining local DWA planning with global A* cluttered environments such as modern warehouses. By relying
on onboard sensor data and locally computed maps, the
system remains resilient to communication failures and scales
efficiently to large teams of robots.
ACKNOWLEDGMENT
The authors would like to express their sincere gratitude
to Dr. Shyam Kamal and his research team for their contin-
uous guidance, insightful feedback, and unwavering support
throughout the course of this project. Their expertise and
mentorship were instrumental in shaping the direction and
quality of this work.
The authors also acknowledge the valuable support provided
by the Drone Lab, IIT (BHU) Varanasi, for granting access to
simulation hardware and resources essential for the experi-
mental validation of the proposed algorithms.
R EFERENCES
[1] P. R. Wurman, R. D’Andrea, and M. Mountz, “Coordinating hundreds of
cooperative, autonomous vehicles in warehouses,” AI Magazine, vol. 29,
no. 1, pp. 9–20, 2008.
[2] B. Reily, T. Mott, and H. Zhang, “Decentralized and communication-
free multi-robot navigation through distributed games,”
https://arxiv.org/abs/2012.09335, 2020.
[3] R. Chandra, V. Zinage, E. Bakolas, P. Stone, and J. Biswas, “Deadlock-
free, safe, and decentralized multi-robot navigation in social mini-games
via discrete-time control barrier functions,” in Proceedings of the IEEE
Conference on Robotics and Automation, 2023.
[4] E. A. Jensen and M. Gini, “Effects of communication restriction on
online multi-robot exploration in bounded environments,” https://www-
users.cse.umn.edu/ gini/publications/papers/Jensen18dars.pdf, 2019.
[5] G. Chen et al., “Distributed non-communicating multi-robot collision
avoidance via map-based deep reinforcement learning,” Sensors, vol. 20,
no. 8, p. 4836, 2020.
[6] Y. Zhang and Others, “Ma-rrdt*: A multi-agent rapidly-exploring ran-
dom disjointed-trees algorithm,” IEEE Transactions on Robotics, vol. 34,
no. 4, pp. 1000–1010, 2018.
[7] M. Alotaibi and A. Al-Rawi, “Multi-robot rapidly exploring random tree
for efficient path planning,” in Proceedings of the IEEE International
Conference on Robotics and Automation (ICRA), 2019.
[8] J. Doe and A. Smith, “Marvel: Multi-agent routing in variable environ-
ments with learning,” IEEE Robotics and Automation Letters, 2020.
[9] K. Lee and L. Chen, “Integrating a* with the dynamic window approach
for robust path planning,” Robotics and Autonomous Systems, 2017.
[10] S. Kim and J. Park, “Multi-module enhanced dwa for navigation in
dense environments,” IEEE Transactions on Automation Science and
Engineering, 2019.
[11] M. Garcia and Y. Liu, “Aco-dwa: An ant colony optimization-based
approach for global-local planning,” in Proceedings of ICRA, 2021.
[12] A. Birk and S. Carpin, “Map merging techniques for multi-robot
systems,” Journal of Field Robotics, 2016.
[13] X. Wang and M. Zhao, “Real-time line-of-sight maintenance in unknown
environments,” in Proceedings of ICRA, 2020.
[14] L. E. Parker, “Distributed robot coordination and control,” in Proceed-
ings of IROS, 2007.
[15] V. Bui et al., “A survey on multi-robot motion planning,” IEEE Trans-
actions on Robotics, 2018.
[16] Q. Wang, J. Li, L. Yang, Z. Yang, P. Li, and G. Xia, “Distributed multi-
mobile robot path planning and obstacle avoidance based on aco–dwa
in unknown complex terrain,” Electronics, vol. 11, no. 14, 2022.
[Online]. Available: https://www.mdpi.com/2079-9292/11/14/2144