RUTGe: Realistic Urban Traffic Generator for Urban Environments
Using Deep Reinforcement Learning and SUMO Simulator
Alberto Baz
´
an Guill
´
en
1 a
, Pablo A. Barbecho Bautista
2 b
and M
´
onica Aguilar Igartua
1 c
1
Networking Engineering Department, Universitat Polit
`
ecnica de Catalunya (UPC), Barcelona, Spain
2
Departamento de El
´
ectrica, Electr
´
onica y Telecomunicaciones, Universidad de Cuenca (UCuenca), Cuenca, Ecuador
Keywords:
SUMO Traffic Generation, Smart Cities, Realistic Simulator, Deep Reinforcement Learning.
Abstract:
We are witnessing a profound shift in societal and political attitudes, driven by the visible consequences of
climate change in urban environments. Urban planners, public transport providers, and traffic managers are
urgently reimagining cities to promote sustainable mobility and expand green spaces for pedestrians, bicycles,
and scooters. To design more sustainable cities, urban planners require realistic simulation tools to optimize
mobility, identify location for car chargers, convert streets to pedestrian zones, and evaluate the impact of
alternative configurations. However, realistic traffic profiles are essential to produce meaningful simulation
results. Addressing this need, we propose a traffic generator based on deep reinforcement learning integrated
with the SUMO simulator. This tool learns to generate an instantaneous number of vehicles throughout the
day, aligning closely with the target profiles observed at the traffic monitoring stations. Our approach gen-
erates accurate 24-hour traffic patterns for any city using minimal statistical data, achieving higher accuracy
compared to existing alternatives. In particular, our proposal demonstrates a highly accurate 24-hour traffic
adjustment, with the generated traffic deviating only by about 5% from the real target traffic. This performance
significantly exceeds that of current SUMO tools like RouteSampler, which struggle to accurately follow the
total daily traffic curve, especially during peak hours when severe traffic congestion occurs.
1 INTRODUCTION
The emergence of Smart Cities has transformed how
urban areas tackle challenges like traffic congestion,
sustainable mobility, and the improvement of public
transportation services. By leveraging advanced tech-
nologies and data analytics, these intelligent cities op-
timize resources and promote more efficient, sustain-
able mobility. In this context, Intelligent Transporta-
tion Systems (ITS) have become essential for man-
aging traffic, reducing emissions, and enhancing user
experiences. However, the effective deployment and
testing of ITS demand accurate simulation environ-
ments capable of replicating real-world urban traffic
patterns.
Mobility hubs, urban planners, traffic engineers,
and public transport providers play a crucial role in
shaping sustainable and livable cities. The reduction
of CO
2
emissions, the enhancement of urban mobil-
a
https://orcid.org/0000-0001-8634-6907
b
https://orcid.org/0000-0002-5281-9208
c
https://orcid.org/0000-0002-6518-888X
ity, and the creation of green spaces, such as the inno-
vative superblocks seen in cities like Barcelona, Paris,
Bremen, and Bergen, require thorough planning and
evaluation. To achieve these goals, realistic simula-
tion tools are essential for testing proposals aimed at
improving urban mobility.
In urban simulation research, open source soft-
ware such as SUMO (Simulation of Urban Mobility)
(Lopez et al., 2018) has become a versatile and scal-
able tool. By offering an open framework, SUMO
allows researchers to access its core modeling com-
ponents, integrate advanced techniques like reinforce-
ment learning, and tailor the system to the specific
needs of individual projects. This makes it a key re-
source for traffic studies requiring the replication of
behaviors in complex urban road networks.
Effective urban simulators not only model road in-
frastructure, but also incorporate dynamic traffic be-
haviors specific to urban settings. SUMO, as a widely
used open-source traffic simulator, provides a flexible
platform for modeling urban traffic. However, their
default tools for generating a 24-hour traffic intensity
pattern might not be sufficiently realistic to evaluate
Bazán Guillén, A., Barbecho Bautista, P. A. and Aguilar Igartua, M.
RUTGe: Realistic Urban Traffic Generator for Urban Environments Using Deep Reinforcement Learning and SUMO Simulator.
DOI: 10.5220/0013375000003941
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 11th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2025), pages 557-564
ISBN: 978-989-758-745-0; ISSN: 2184-495X
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
557

the performance of proposed mobility strategies un-
der varying traffic conditions. Real-world scenarios
involve dynamic traffic patterns influenced by week-
days, weekends, holidays, and variations across the
24-hour cycle. A simple traffic profile fails to capture
these complexities, limiting the usefulness of simula-
tions for practical applications.
To contribute to improving the generation pro-
cess of urban traffic, this paper proposes a method-
ology to automatically generate realistic traffic pat-
terns for any city, using minimal statistical data (e.g.,
average traffic intensity over a 24-hour period). We
have developed and implemented a SUMO-based tool
named RUTGe (Realistic Urban Traffic Generator)
that leverages Deep Reinforcement Learning (DRL)
to learn how to generate the instantaneous number of
vehicles throughout the day, aligning closely with the
target profiles observed at traffic monitoring stations.
Our proposal delivers more accurate and faster results
compared to other methods developed for SUMO to
generate variable 24-hour traffic profiles.
The rest of the paper is organized as follows. Sec-
tion 2 outlines the main works related to our study.
Section 3 presents the fundamentals of our proposed
framework and describes the main algorithms devel-
oped. In Section 4, we discuss simulation results. Fi-
nally, Section 5 concludes the paper and suggests di-
rections for future research.
2 RELATED WORK
Simulation tools like SUMMIT (Cai et al., 2020)
exemplify how urban simulations integrate road in-
frastructure and dynamic traffic behaviors unique
to urban contexts, having been employed to assess
autonomous driving systems in densely populated
and unregulated traffic environments where vehicle-
pedestrian interactions might be highly complex.
Similarly, SceneGen (Tan et al., 2021) has demon-
strated the effectiveness of auto-regressive neural net-
works in generating realistic traffic scenarios with-
out relying on predefined rule-based heuristics. This
approach is particularly valuable for modeling self-
driving vehicles, by addressing traditional challenges
in capturing the complexities of urban traffic.
Additionally, the integration of real-world data
has proven crucial for advancing urban traffic simula-
tions. For instance, a Markov-chain traffic model was
utilized in (Arias et al., 2017) to predict electric vehi-
cle (EV) charging demand at fast-charging stations in
urban areas. This model incorporates real-time traffic
data obtained from closed-circuit television (CCTV)
cameras in Seoul, South Korea, to develop an EV
charging-power demand framework. This approach
highlights spatial and temporal traffic patterns and the
value of context-aware simulations, emphasizing the
need for adaptable, data-driven tools tailored to urban
environments.
The integration of human preferences and behav-
ior has also been pivotal in advancing traffic simu-
lations. The proposal (Cao et al., 2024) combines
reinforcement learning with human feedback to train
agents that not only replicate realistic human behav-
ior but also comply with traffic regulations, thereby
enhancing the realism of existing traffic models. This
approach underscores the significance of incorporat-
ing human judgment into the design of traffic simula-
tions to more accurately mirror real-world conditions.
The approach (Padr
´
on et al., 2023) leverages real-
world data by developing a traffic model that utilizes
real-time data from induction loop detectors installed
throughout the city. This model predicts traffic flow
and generates realistic Origin-Destination (OD) traf-
fic matrices, achieving more accurate route lengths
and a better distribution of traffic sources and desti-
nations compared to the dfrouter tool available in
SUMO (Lopez et al., 2018).
All these studies collectively highlight the impor-
tance of integrating data-driven and human-centered
approaches to develop realistic urban traffic simula-
tions. Building on this foundation, our work takes a
step further by addressing the challenge of generat-
ing a realistic urban traffic profile over a complete 24-
hour period corresponding to a typical working day.
To achieve this, we have developed a Deep Reinforce-
ment Learning model capable of accurately capturing
and replicating the dynamic traffic patterns of a city
over the course of an entire day.
3 PROPOSAL TO GENERATE
REALISTIC SUMO TRAFFIC
SUMO includes a tool called dfrouter (Lopez et al.,
2018), designed to generate traffic based on data col-
lected from detection points, such as induction loops.
However, this tool somehow fails when applied to ur-
ban areas, as it was originally intended for highway
scenarios. An alternative approach in (Padr
´
on et al.,
2023) improves dfrouter’s performance in urban en-
vironments, specifically Valencia, Spain, by gener-
ating realistic traffic with route lengths suited to the
characteristics of the city.
However, to the best of our knowledge, no ex-
isting solution aims to generate realistic traffic for a
full day for any city solely based on its map and pre-
collected traffic data while leveraging reinforcement
VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems
558

learning. Our proposal aims to develop a reinforce-
ment learning-based approach using the SUMO sim-
ulator to create a realistic traffic model capable of pro-
ducing the desired average hourly vehicle flow for a
given city. Moreover, our DRL model can be seam-
lessly extended to cover 24-hour periods, enabling the
generation of traffic patterns that not only align with
historical averages from traffic metering stations, such
as induction loops, but also adapt to specific scenar-
ios, including holidays, weekends, and peak hours.
Our objective is to generate realistic 24-hour
traffic patterns representative of an average day in
any given city. To achieve this, the process be-
gins with a setup phase that involves preparing the
simulation environment. This includes selecting the
city map and obtaining real-world traffic measure-
ments for analysis. For our case study, we fo-
cused on Barcelona, Spain, using OpenStreetMap
(OSM) (OpenStreetMap, 2024) data. The map was
converted into a SUMO-compatible format using
SUMO’s netconvert tool. Historical traffic inten-
sity data, measured in number of vehicles per hour at
various traffic monitoring stations across the city, was
provided by the Barcelona City Council. From this
dataset, we selected data from five key monitoring sta-
tions. These locations were replicated in our SUMO
simulation scenario by placing induction loops at
the same positions as the real-world sensors that col-
lected the historical data.
After the configuration phase, the next step in-
volved developing a realistic traffic generator tailored
to the specific scenario. This traffic-generating agent
is based on a reinforcement learning model designed
to create traffic patterns that closely align with a pre-
defined target traffic profile for a given hour. Fol-
lowing a necessary training period, the model can ac-
curately generate the desired traffic for the specified
hour, establishing the agent as a key component of
the realistic traffic simulation system for the city.
Once the traffic generation model is trained, it will
be executed in SUMO for each of the 24 hours to
simulate a complete daily traffic profile. The model
generates the required traffic for each hour while con-
sidering residual traffic from preceding hours, which
may influence subsequent traffic conditions. Upon
completing the 24-hour simulation cycle, the system
exports a file containing vehicle data and routes that
replicate the desired traffic intensity profile for the
city on a typical working day, as it represents the most
representative and congested traffic conditions.
3.1 Deep Reinforcement Learning for
Realistic Traffic Simulation
We have developed a model-free Deep Reinforcement
Learning approach to generate realistic urban traffic
patterns. SUMO is used as the simulation environ-
ment, acting as a black-box traffic model that pro-
vides state observations as output after simulating one
hour of traffic. The RL agent’s objective is to learn a
policy that dynamically adjusts the number of vehi-
cles introduced into the network during the simula-
tion, minimizing the deviation between the generated
traffic and the target traffic. This section outlines the
formalization of the RL framework, its components,
and the implementation methodology.
3.1.1 Model-Free Reinforcement Learning
Framework, 1-Hour Agent
Reinforcement learning is a paradigm in which an
agent interacts with its environment to learn behavior
that maximizes a cumulative reward. In the model-
free approach, the agent does not attempt to construct
or predict the transition dynamics of the environment.
Instead, it learns a policy directly based on observed
states, actions, and rewards.
The proposed RL system is framed as a Markov
Decision Process (MDP), defined by the tuple
(S, A, R, P), where S is the set of states s observed
from the environment, A is the set of actions a the
agent can take, R(s,a) is the reward function pro-
viding feedback for taking action a in state s, and
P(s
′
|s, a) is the transition probability function (from
state s to state s
′
), which is implicit in the model-free
RL and learned indirectly.
In our implementation, SUMO served as the envi-
ronment, and its outputs after one hour of simulation
provided the observations necessary to define the state
and calculate the reward. The RL framework was in-
stantiated with the following four components:
a. Environment: SUMO Simulator
SUMO is widely used as a high-fidelity simulation
environment for urban traffic. In our simulation sce-
nario, after each simulation step (representing one ex-
ecution during the simulated hour), SUMO generates
traffic metrics such as vehicle intensity (vehicles/h),
congestion level (vehicles/km
2
), and average speed
(km/h), among others. In our case, vehicle inten-
sity is used to compute the current observation of the
environment and subsequently calculate the reward
R(w, a) for the selected action.
b. State, s
The state in the reinforcement learning framework en-
capsulates the traffic injection configuration for the
RUTGe: Realistic Urban Traffic Generator for Urban Environments Using Deep Reinforcement Learning and SUMO Simulator
559

simulation. Specifically, the state consists of the num-
ber of vehicles to be introduced into the traffic net-
work, distributed throughout the hour being analyzed
in the SUMO simulation environment.
The injection of traffic, specifically the Origin-
Destination (OD) trips, plays a crucial role in defining
the traffic distribution within the simulation environ-
ment. The number of vehicles to be introduced into
the network determines the number of OD pairs that
will be distributed across the map. These OD pairs
are used to generate trips via the OD2Trips tool in
SUMO, which converts OD pairs into individual trips
within the network.
To ensure a realistic traffic flow that aligns with
the real-world distribution of vehicles, trips are dis-
tributed across the network using the via attribute,
which allows specifying the exact routes that vehi-
cles will follow. Subsequently, the duarouter tool
in SUMO is used to generate the routes for each trip.
These routes represent the paths that vehicles will take
during the simulation. The selected tools were in-
spired by the work (Barbecho Bautista et al., 2022).
Once the trips and their respective routes are gen-
erated, they form the basis for running the simulation
in SUMO. The simulation outputs consist of intensity
measurements recorded by induction loops placed on
the map throughout the simulated hour. These out-
puts serve as the observed states in the reinforcement
learning framework and play a crucial role in evaluat-
ing the agent’s performance. Additionally, they guide
policy updates during training, ensuring that the sim-
ulated traffic conditions accurately reflect real-world
traffic patterns.
To improve the efficiency and stability of rein-
forcement learning, the state s was normalized to [0, 1]
using the minimum and maximum values observed
for each traffic metric. This ensured consistent scaling
and prevented any feature from dominating the learn-
ing process.
c. Action, a
Actions are defined as the number of vehicles injected
into the traffic network at the start of each simulation
step. The agent learns to determine the optimal in-
jection rate to achieve the desired traffic conditions.
The action A in the proposed reinforcement learning
framework modifies the current state S to produce a
new state S
′
. This modification is defined as:
S
′
= S · A, (1)
where A is a continuous scalar value within the range
[−0.5, 2.0]. This design allows the agent to dynami-
cally scale the current state, enabling reductions (A <
1) and amplifications (A > 1) of the traffic metrics.
After applying the action, the new state S
′
is eval-
uated to compute the reward. This reward, along with
S
′
, the action A, and the original state S, is stored in
the experience buffer for training the Proximal Policy
Optimization (PPO) algorithm.
By iteratively adjusting the state through actions,
the agent learns to select optimal scaling factors (A)
that minimize the deviation from the desired traffic
conditions.
d. Reward, R(s, a)
The reward is calculated as the negative mean squared
error (MSE) between the target traffic metrics and
the corresponding traffic output generated by SUMO.
The use of a negative value ensures that the reinforce-
ment learning agent interprets smaller deviations as
higher rewards, aligning with the objective of mini-
mizing the error. The reward function is formally de-
fined as:
R(s, a) = −
1
N
N
∑
i=1
(y
i
− ˆy
i
)
2
, (2)
where y
i
represents the target value for a traffic met-
ric and ˆy
i
is the observed SUMO value for a given
state s and action a. This reward helps the agent min-
imize the deviation between the current traffic condi-
tions and the desired traffic pattern.
The SUMO simulator outputs ( ˆy
i
) the average traf-
fic intensity values recorded by the virtual detectors
placed within the simulated map. These detectors are
placed at specific locations that correspond to real-
world traffic detectors, allowing direct comparisons
between actual and simulated traffic conditions.
The target values (y
i
) represent the average hourly
traffic intensities recorded during a typical weekday,
provided by the Barcelona City Council. These tar-
gets are based on real-world data collected from de-
tectors located at the same positions as the virtual
ones in the simulation.
To ensure adaptability, the reinforcement learning
model is trained using a range of target values rep-
resenting various traffic scenarios. This approach en-
ables the model to generalize and adjust the simulated
traffic for any future demand. By minimizing the de-
viation between the simulated and target intensities,
the model learns to dynamically regulate traffic injec-
tion rates to match desired patterns within a one-hour
simulation period.
The target intensities are integrated into the re-
ward function as described in Eq. (2), where the neg-
ative mean squared error (MSE) between the SUMO
outputs and the targets guides the agent to improve
simulation accuracy and realism.
Additional Reward Components. To enhance the
learning process, the reward function incorporates the
VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems
560

following elements:
• Step Penalty: At every step, the agent receives a
small negative reward, R
step
= −λ, where λ > 0.
This penalty encourages the agent to find solu-
tions in fewer steps, promoting efficiency.
• Goal Achievement Reward: If the agent
achieves an MSE smaller than a predefined thresh-
old (MSE < 10
−5
), it is granted a large positive
reward, R
goal
= +η (η ≫ 1). This reward strongly
reinforces successful policies that achieve the ob-
jective with high precision.
Complete Reward Function. The complete reward
at each step, R
total
(s, a), integrates these components
and is defined as follows:
R
total
(s, a) =
(
R(s, a) − λ, if MSE ≥ 10
−5
+η, if MSE < 10
−5
(3)
To balance the incentive for achieving the objec-
tive with the penalty for prolonged episodes, the step
penalty and the goal achievement reward were defined
as follows:
• Step Penalty (λ): A value of λ = 0.01 was used to
lightly penalize each step, encouraging the agent
to seek efficient solutions without overshadowing
the primary objective.
• Goal Achievement Reward (η): A large posi-
tive reward of η = 10.0 was granted when the
agent achieved a mean squared error smaller than
0.00001. This strong positive signal prioritized
policies that met the desired precision.
These values ensured a balance between promot-
ing efficient behavior and strongly reinforcing suc-
cessful outcomes.
3.1.2 Policy, Step Function, and Training
Process
The agent’s policy is trained using a model-free rein-
forcement learning framework based on the Proximal
Policy Optimization (PPO) algorithm. PPO is well-
suited for tasks involving complex environments such
as SUMO due to its stability and efficiency. The agent
iteratively updates its policy to maximize cumulative
rewards through controlled updates, ensuring effec-
tive exploration and exploitation.
A custom step function integrates SUMO with the
RL framework, facilitating interaction between the
agent and the simulation environment. For each step:
• The agent takes an action a, modifying the vehicle
injection rate in SUMO.
• SUMO simulates one hour of traffic and provides
updated metrics that define the next state s
′
.
• The reward R(s, a) is calculated using the mean
squared error (MSE) between the target and simu-
lated traffic metrics and also taking the additional
reward components, as defined in Eq. (3).
The training process consists of multiple episodes,
each comprising several one-hour simulations, each
time with different target values. During each
episode:
• The agent observes the current state s and calls
SUMO to obtain the observation from the SUMO
outputs.
• The step function calculates the reward and tran-
sitions the system to the next state s
′
.
• The PPO algorithm updates the policy based on
cumulative rewards, enabling the agent to mini-
mize the deviation from the target traffic metrics.
This iterative process ensures that the agent learns
a robust policy for dynamically adjusting traffic con-
ditions to achieve realistic and efficient simulations.
Additionally, using different target values in each
simulation helps develop an agent capable of adapt-
ing to a wide range of targets.
3.2 24-Hours Traffic Generator
Once the 1-hour traffic model is trained, it is utilized
to generate the desired traffic pattern over a 24-hour
period. For each hour, the 1-hour DRL-based model
described in Sec. 3.1 is executed with the target traf-
fic intensity for that specific hour, as outlined in Al-
gorithm 2. At the end of the simulation, the residual
traffic is collected. This residual traffic represents ve-
hicles remaining on the network after injecting the de-
sired number of vehicles at the start of the given hour.
These delayed vehicles contribute to the traffic inten-
sity of subsequent hours, serving as additional input
for the following simulations.
The residual traffic is accumulated and subtracted
from the target value for the subsequent hour, under
the assumption that the carried-over traffic from the
previous hour will already be present. This approach
fixing the traffic for one hour, accounting for resid-
ual traffic from the previous hour, and subtracting it
from the target for the next hour is applied iteratively
throughout the 24-hour period, as detailed in Algo-
rithm 1. This process ensures the generation of a traf-
fic profile that closely approximates the desired pat-
tern for the entire day.
As a result, a set of vehicles and routes is gen-
erated with varying departure times and locations,
RUTGe: Realistic Urban Traffic Generator for Urban Environments Using Deep Reinforcement Learning and SUMO Simulator
561

Algorithm 1: 24h-TGA ⇒ 24-hours Traffic Generation Al-
gorithm.
Data: Map file (.net.xml), start time T
start
,
end time T
end
, hourly target intensities
target[]
Result: Complete route file
(final routes.rou.xml)
1 Initialize final routes.rou.xml as an
empty file;
2 for i ← T
start
to T
end
do
// Generate traffic for hour i
3 [routes.rou.xml, residual[]] =
1h-TGA(i, target[i]) ← Alg. 2;
4 Append routes.rou.xml to
final routes.rou.xml;
// Adjust targets for subsequent
hours
5 for j ← i + 1 to T
end
do
6 target[ j] = target[ j] -
residual[ j];
7 return final routes.rou.xml;
Algorithm 2: 1h-TGA ⇒ 1-hour Traffic Generation Algo-
rithm.
Data: Hour i, target intensity target[i]
Result: Route file (routes.rou.xml), traffic
residual vector (residual[])
1 Initialize routes.rou.xml as an empty file;
2 Initialize residual[] as a vector of zeros;
3 while MSE(observed intensity, target[i])
> threshold do
4 Call PPO model with target[i], state S
to propose action a and adjust traffic
injection;
5 Update observed traffic intensity and
calculate R(s, a);
6 Make S
′
= S;
7 Compute residual[] as the traffic delayed
to subsequent hours;
8 return routes.rou.xml, residual[];
populating the map with vehicles that undertake trips
throughout the simulation. This method achieves the
desired average hourly traffic intensity, enabling the
creation of realistic, time-dependent traffic flows that
reflect real-world patterns and maintain consistency
over the 24-hour simulation period.
4 PERFORMANCE EVALUATION
The proposed tool demonstrates remarkable versatil-
ity, allowing the generation of realistic traffic patterns
for any city map available in OpenStreetMaps (Open-
StreetMap, 2024). By integrating traffic intensity data
for specific detection points in the city, the tool pro-
duces simulated traffic that matches the average traffic
intensity observed at these points. This capability en-
ables users to replicate realistic traffic conditions ac-
cording to their needs, making the tool highly adapt-
able for diverse simulation scenarios. To demonstrate
our methodology, we used the Barcelona simulation
scenario provided by OpenStreetMap (OSM) (Open-
StreetMap, 2024), depicted in Fig. 1. In our study,
we selected five out of the fifteen available traffic sta-
tions in Barcelona to demonstrate the methodology
and evaluate the accuracy obtained.
Figure 1: Simulation scenario of a 45 km
2
area in Barcelona
(9 km x 5 km), with traffic intensity monitored at five key
traffic stations (indicated by black icons). Map sourced
from OSM (OpenStreetMap, 2024).
The RUTGe tool leverages our pre-trained DRL-
based model, which requires an initial investment of
near 5 hours in training time, with a computer with
these features: CPU i9-12900K, 3.2 Ghz, 30MB;
RAM 32GB DDR5 4800Mhz. Once trained, our
RUTGe tool efficiently generates traffic for the spec-
ified time range, averaging just 15 minutes per sim-
ulated 24 hours of real-world traffic. Users can cus-
tomize the time range and adjust traffic intensity tar-
gets for each hour to meet specific traffic profiles.
This underscores the tool’s efficiency, showing that
the initial investment in training time is well worth
the valuable outcomes it delivers. Note that current
SUMO tools, such as RouteSampler, require around 1
hour 15 min to simulate 24 hours of traffic.
4.1 Simulation Results
The progression of the reward over iterations for a sin-
gle hour is shown in Fig. 2. The selected hour (9:00
VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems
562

AM) was the peak traffic hour during the day. We ob-
serve that the reward achieved during each simulation
step consistently improves as the model iteratively re-
fines its policy to match the desired traffic intensity.
The graph shows a steady increase in reward until it
converges to the maximum value, indicating that the
simulated traffic matches the target intensity.
Figure 2: Reward progression during traffic generation for
a single hour (9:00 AM).
To validate the accuracy of the tool, simulations
were performed for a 24-hour urban scenario. Con-
fidence intervals of 99%, computed from 10 repeti-
tions with different seeds, are shown. These intervals
are very small, indicating consistency in the generated
models. Figs. 3a and 3b compare the simulated traf-
fic intensities (dashed red lines) with real-world data
(solid blue lines) for each hour. Results indicate that
the traffic generated by our RUTGe tool aligns more
closely with the target traffic intensity compared to
RouteSampler, highlighting the superior precision of
our traffic generation tool. The average relative error
is 5.5% with RUTGe and 8.5% with RouteSampler.
The main difference is that our approach takes into
account the residual traffic from each hour, which im-
pacts subsequent hours (see Sec. 3.2), whereas Route-
Sampler does not.
Figure 4a presents the spatial traffic distribu-
tion visualized on the colormap generated using the
SUMO tool plot net dump.py for a 24-hour simula-
tion. The generated traffic aligns well with real traffic
patterns on the most congested main roads, while also
being evenly distributed across the remaining road
networks, effectively reflecting realistic traffic behav-
ior. This result underscores the ability of the RUTGe
tool to model complex urban traffic systems with high
fidelity. Fig. 4b shows that current SUMO tools, such
as RouteSampler, tend to concentrate the majority of
traffic around traffic stations to meet the target traffic
intensity. However, the traffic distribution across the
rest of the scenario is less evenly spread.
5 CONCLUSIONS AND FUTURE
WORK
This work presents the development of a tool for gen-
erating realistic urban traffic patterns using deep rein-
forcement learning. By utilizing a city map extracted
from OpenStreetMap and historical traffic intensity
data, the tool trains a reinforcement learning agent ca-
pable of achieving the desired average traffic intensity
for a one-hour simulation. Once trained, the model
can extend its functionality to simulate multiple hours
(e.g., a day) with varying traffic intensity targets, aim-
ing to generate traffic patterns that closely resemble
real-world conditions.
For validation, the tool was applied to a 45 km
2
section of Barcelona to simulate 24 hours of traffic
on an average working day. The results demonstrate
promising accuracy, achieving a relative error of 5.5%
compared to real-world traffic intensity data. Addi-
tionally, the short average execution time of 15.2 min-
utes per simulation justifies the initial training time of
5 hours, as the trained model can be reused to simulate
any combination of hours and intensity targets within
the same environment. Furthermore, the traffic distri-
bution achieved on the map aligns well with realistic
urban patterns, making the tool valuable for simulat-
ing various services and incidents in urban environ-
ments. The generated traffic patterns can serve as the
foundation for advanced studies in intelligent trans-
portation systems, urban planning, and autonomous
vehicle testing.
As future work, we plan to develop a Federated
Deep Reinforcement Learning scheme, where each
independent traffic station will locally train its own
model. Stations will share hyperparameters, collabo-
ratively constructing a global prediction model. This
approach is expected to significantly reduce training
time while maintaining model accuracy.
ACKNOWLEDGEMENTS
This work was partially supported by the Span-
ish Government under these research projects
funded by MCIN/AEI/10.13039/501100011033:
DISCOVERY PID2023-148716OB-C32; MOBI-
LYTICS TED2021-129782B-I00 (also funded by
the European Union NextGenerationEU/PRTR);
COMPROMISE PID2020-113795RB-C31; predoc-
toral scholarship associated with the ”Generaci
´
on
de Conocimiento” Projects, Call 2022, PRE2021-
099830. Also, by the Generalitat de Catalunya
AGAUR grant ”2021 SGR 01413”.
RUTGe: Realistic Urban Traffic Generator for Urban Environments Using Deep Reinforcement Learning and SUMO Simulator
563

(a) 24-hours traffic generated using our proposal RUTGe. (b) 24-hours traffic generated using the RouteSampler tool.
Figure 3: Comparison of simulated traffic intensities with real-world data (target traffic to adjust the model). 99% confidence
intervals are shown in the RUTGe model and RouteSampler tool. The average relative error is 5.5% and 8.5%, respectively.
The simulation time is 15 min and 19 min, respectively.
(a) SUMO colormap using our proposal RUTGe. (b) SUMO colormap using the RouteSampler tool.
Figure 4: Traffic distribution across the map for a 24-hour simulation. Urban scenario depicted in Fig. 1.
REFERENCES
Arias, M. B., Kim, M., and Bae, S. (2017). Prediction
of electric vehicle charging-power demand in realistic
urban traffic networks. Applied energy, 195:738–753.
Barbecho Bautista, P., Urquiza Aguiar, L., and Aguilar Igar-
tua, M. (2022). How does the traffic behavior change
by using sumo traffic generation tools. Computer
Communications, 181:1–13.
Cai, P., Lee, Y., Luo, Y., and Hsu, D. (2020). SUMMIT: A
simulator for urban driving in massive mixed traffic.
In 2020 IEEE International Conference on Robotics
and Automation (ICRA), pages 4023–4029. IEEE.
Cao, Y., Ivanovic, B., Xiao, C., and Pavone, M. (2024). Re-
inforcement learning with human feedback for realis-
tic traffic simulation. In 2024 IEEE International Con-
ference on Robotics and Automation (ICRA), pages
14428–14434. IEEE.
Lopez, P. A., Behrisch, M., Bieker-Walz, L., Erdmann, J.,
Fl
¨
otter
¨
od, Y.-P., Hilbrich, R., L
¨
ucken, L., Rummel, J.,
Wagner, P., and Wiessner, E. (2018). Microscopic
Traffic Simulation using SUMO. In 2018 21st In-
ternational Conference on Intelligent Transportation
Systems (ITSC), pages 2575–2582.
OpenStreetMap (2024). https://www.openstreetmap.org.
Padr
´
on, J. D., Hern
´
andez-Orallo, E., Calafate, C. T., Soler,
D., Cano, J.-C., and Manzoni, P. (2023). Realistic traf-
fic model for urban environments based on induction
loop data. Simulation Modelling Practice and Theory,
125:102742.
Tan, S., Wong, K., Wang, S., Manivasagam, S., Ren, M.,
and Urtasun, R. (2021). SceneGen: Learning to Gen-
erate Realistic Traffic Scenes . In Proceedings of the
IEEE/CVF Conference on Computer Vision and Pat-
tern Recognition, pages 892–901.
VEHITS 2025 - 11th International Conference on Vehicle Technology and Intelligent Transport Systems
564