CrowdSim++: Unifying Crowd Navigation and Obstacle Avoidance

Marco Rosano

, Danilo Leocata

, Antonino Furnari

1,2

and Giovanni Maria Farinella

1,2

FPV@IPLAB, Department of Mathematics and Computer Science, University of Catania, Catania, Italy

Next Vision s.r.l., Catania, Italy

Keywords:

Crowd Navigation, Reinforcement Learning, Obstacle Avoidance, Crowd Simulator.

Abstract:

In recent years, signiﬁcant advancements in learning-based technologies have propelled the development of

autonomous robotic systems designed to assist humans in challenging scenarios during their daily activi-

ties. This research focuses on enhancing robotic perception and control, particularly in navigating complex,

crowded environments. Traditional approaches often treat static and dynamic components separately, limiting

the robots’ real-world performance. We propose CrowdSim++, an extension of the open-source CrowdSim

simulator (Chen et al., 2019), to unify crowd navigation and obstacle avoidance. CrowdSim++ enables training

navigation policies in dynamically generated environments or real-world ﬂoor plans, using a 2D lidar sensor

and a “person sensor” for enhanced perception. Our experiments demonstrate that Reinforcement Learning-

based navigation policies trained in complex environments with humans outperform those trained in simpler

scenarios. Additionally, providing robots with specialized sensors to accurately distinguish between static and

dynamic obstacles is essential for achieving superior performance. To advance research in autonomous navi-

gation, the source code and dataset of realistic ﬂoor plans are available at the following link.

1 INTRODUCTION

In recent years, there has been signiﬁcant progress

in developing learning-based technologies for robotic

systems, aiming to create autonomous agents capa-

ble of assisting humans in their daily activities (Chen

et al., 2013; Miller, 2006). Research in this ﬁeld has

particularly focused on various sub-problems related

to the development of complex robotic agents, includ-

ing perception for action through algorithms process-

ing data acquired via cameras and other sensors (Chen

et al., 2021), mechanical arm control (Dong and

Zhang, 2023), and navigation (Otte, 2015). Notably,

considerable efforts have been made to combine per-

ception and control to perform complex operations

across different environments (Yaar et al., 2024).

Several learning-based solutions have emerged to

address the autonomous navigation problem, enabling

robots to plan and execute routes to destinations with-

out relying on environmental maps (Savva et al.,

2019) or even amidst moving people in crowded set-

tings (Chen et al., 2019). These navigation models

are typically trained in simulated environments (Otte,

2015), providing a wide variety of navigation sce-

narios essential for learning optimal navigation poli-

cies, avoiding the difﬁculties of learning from real ob-

servations. Indeed, collecting this experience in the

real world would be impractical due to the time re-

quired and the fragility and cost of the robot’s hard-

ware (Rosano et al., 2023).

Navigating crowded environments is of particular

interest as it represents a fundamental requirement for

designing robotic systems that operate alongside hu-

mans in real environments (M

oller et al., 2021). De-

spite the signiﬁcant progress in autonomous naviga-

tion, the design of systems for crowded environments

has often approached the dynamic component (mov-

ing people) separately from the static component (en-

vironment layout) (Chen et al., 2017).

While previous works have addressed point-goal

robot navigation and crowded navigation as separate

tasks (Mavrogiannis et al., 2023), we postulate that a

robot can beneﬁt from learning to perform both tasks

simultaneously. This approach leverages the inter-

actions between humans and the environment (e.g.,

humans won’t collide with walls), leading to more

effective navigation strategies. Speciﬁcally, we pro-

pose to train a navigation policy in simulated envi-

ronments that mirror the complexity of real-world set-

tings, providing the perceptual capabilities necessary

to address the navigation challenges. To support our

investigation, we propose CrowdSim++, which ex-

Rosano, M., Leocata, D., Furnari, A. and Farinella, G. M.

CrowdSim++: Unifying Crowd Navigation and Obstacle Avoidance.

DOI: 10.5220/0013192300003890

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

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 2, pages 533-542

ISBN: 978-989-758-737-5; ISSN: 2184-433X

533

Navigation episodes in randomly generated environments with 3 obstacles

Humans

trajectories

Human

Robot

without lidar

Robots

trajectories

Robot

with lidar

Obstacle

Humans

trajectories

Human

Robot

without lidar

Robots

trajectories

Robot

with lidar

Obstacle

Figure 1: The proposed CrowdSim++ simulator enables the training of policies for crowd navigation in complex environments

with both obstacles and moving humans. The robot equipped only with the “person sensor” (yellow circle, gray trajectory)

lacks awareness of the environment’s layout and exhibits limited performance, even when trained in contexts similar to the test

scenario. In contrast, the robot equipped with both the “person sensor” and lidar sensor (blue circle, red trajectory) successfully

avoids collisions with static and dynamic obstacles and reaches the goal effectively. The ﬁgure represents snapshots of the

concluded navigation episodes.

tends the functionality of the CrowdSim open-source

simulator (Chen et al., 2019) to set crowd navigation

in randomly generated complex environments or in

environments derived from real-world ﬂoor plans. In

the ﬁrst case, the environments are generated with a

variable number of obstacles of different sizes to cre-

ate scenarios with various levels of complexity. In the

second case, we reproduce ﬂoor plans from real envi-

ronments using the popular Gibson dataset (Xia et al.,

2018), enabling realistic navigation scenarios. To en-

able the robot’s perception of its surroundings, we im-

plemented a 2D lidar sensor that provides informa-

tion on the distance of obstacles in the environment.

This setup mirrors realistic scenarios where a robot

is equipped with lidar to support navigation. Figure 1

shows examples of navigation episodes performed us-

ing the proposed simulator.

To validate our claims and show the usefulness of

CrowdSim++, we train and evaluate different naviga-

tion policies based on Reinforcement Learning (Chen

et al., 2019; Chen et al., 2017) in various complex sce-

narios and compare their performance with the same

baselines trained in simpler environments. Our results

demonstrate that navigation models signiﬁcantly ben-

eﬁt from training in more challenging settings. Addi-

tionally, equipping robots with proper perception abil-

ities that can clearly observe and differentiate between

static and dynamic obstacles is crucial for achieving

superior performance.

The contributions of this work are as follows:

• we propose CrowdSim++, an extended version

of the CrowdSim simulator (Chen et al., 2019)

for training learning-based navigation models in

crowded environments with complex geometry.

The simulator supports the dynamic generation of

environments or the import of ﬂoor plans from

real environments and, in addition to the “human

sensor” originally included in CrowdSim (Chen

et al., 2019), includes the use of a lidar sensor,

necessary for the robot to perceive its surround-

ings;

• we demonstrate that navigation policies trained in

simpler environments exhibit limited generaliza-

tion ability when applied to more complex lay-

outs. Through experiments, we show that training

navigation models in complex environments with

moving humans is essential for developing effec-

tive navigation policies.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

534

• we publicly release the simulator’s source code

along with a ﬂoor plan dataset of real-world en-

vironments to facilitate the development of ad-

vanced navigation systems for complex, crowd-

ﬁlled scenarios.

2 RELATED WORK

Crowd Navigation Systems. In the context of au-

tonomous navigation in crowded environments, var-

ious systems have been proposed over the years to

manage the dynamic nature of space conﬁgurations.

A common approach involves considering pedestrians

with simple kinematics, moving at a constant speed

along mostly straight trajectories. Reactive methods

adopt this assumption and make decisions based on

one-step interaction rules to avoid collisions in the im-

mediate future (Van den Berg et al., 2008). However,

they fail to model human behavior or the future evo-

lution of the environment, often resulting in unnatural

behaviors and the “freezing robot” problem (P

erez-

D’Arpino et al., 2021), where the planner halts and

repeatedly replans the actions to be taken. To ad-

dress these limitations, other methods aim to under-

stand human intentions in terms of trajectory and des-

tination, to capture a long-term vision of the environ-

ment’s evolution (Kuderer et al., 2012).

Recent advancements in deep learning have sig-

niﬁcantly impacted the ﬁeld of robot navigation (Otte,

2015). This new approach facilitates learning nav-

igation policies from demonstrations of desired be-

havior in an end-to-end manner, starting from the

robot’s raw observations. Imitation Learning and Be-

havior Cloning strategies leverage expert supervision

to train deep models in a supervised manner (Bo-

jarski et al., 2016). Nevertheless, these systems of-

ten struggle to generalize beyond the scenarios en-

countered during training. To overcome the prob-

lem, new systems based on Reinforcement Learning

(RL) have emerged, capable of autonomously collect-

ing the necessary experience within simulated envi-

ronments (Rosano et al., 2021; Zhu and Zhang, 2021).

These systems have proven effective for learning from

large amounts of simulated data (Savva et al., 2019).

Nevertheless, these methods are often trained in static

environments without moving people. Concurrent re-

search has focused on developing methods for nav-

igation in crowded environments, utilizing simula-

tion systems that can mimic human movement, al-

beit with simplistic dynamics (P

erez-D’Arpino et al.,

2021). For instance, Chen et al. (Chen et al., 2017)

proposed an RL-based model that encodes the esti-

mated time to the goal given the positions and veloc-

ities of neighbors. Meanwhile, the authors of (Chen

et al., 2019) introduced a method to capture crowd-

robot interactions in the presence of many humans

using a self-attention system to estimate their inﬂu-

ence on future states. Unfortunately, most of these

approaches involve humans isolated from the environ-

mental context, and only a limited number of methods

have addressed the problem using environments with

more complex geometries (Liu et al., 2020). In this

work, we demonstrate that navigation policies trained

in simpler struggle when applied to more complex

layouts, and that training navigation models in real-

istic contexts—featuring both static and dynamic ob-

stacles—can be advantageous for learning advanced

navigation policies, leveraging the capabilities of the

proposed simulator.

Simulators and Datasets for Navigation Systems.

Simulators are fundamental tools for developing

robotic navigation systems. Historically, due to the

scarcity of advanced simulation platforms, many re-

searchers used simple, ad-hoc simulators (Chen et al.,

2017) or commercial products (Staranowicz and Mar-

iottini, 2011). More recently, advanced tools capable

of simulating human presence and human-robot inter-

actions have been developed and made freely avail-

able to the scientiﬁc community (M

oller et al., 2021).

The usability of these simulators has been greatly en-

hanced by their integration with robotic middleware

that provides a uniﬁed environment for using vari-

ous robot software (Quigley et al., 2009) . The rise

of deep learning systems and the success of Rein-

forcement Learning methods in robotics have spurred

a rapid adoption of simulation systems capable of

handling large-scale datasets (Zhu and Zhang, 2021).

One area of focus has been the development of pho-

torealistic simulators (Savva et al., 2019; Xia et al.,

2018), which utilize manually reconstructed indoor

environments (Khanna et al., 2024) or those captured

using 3D scanners (Xia et al., 2018). In this context,

the ﬁdelity of virtual images-ensuring they accurately

reﬂect the characteristics of real-world images-is cru-

cial for applying the navigation system in real envi-

ronments. Concurrently, efforts have concentrated on

simulators designed for crowd navigation, capable of

realistically simulating human movement within re-

constructed environments (Chen et al., 2019; Kuderer

et al., 2012). These simulators typically employ high-

level sensors, thus they focused more on emulating

human movement than on visual rendering (Van den

Berg et al., 2008; Chen et al., 2019). However, only

few studies have explored joint scenarios involving

navigation in complex environments alongside human

presence (Liu et al., 2020). In contrast to previous

studies (Cancelli et al., 2023; P

erez-D’Arpino et al.,

CrowdSim++: Unifying Crowd Navigation and Obstacle Avoidance

535

Adairsville (∼5.5 x 10 meters)

Bonesteel (∼ 10 x 9 meters)

Highspire (∼ 6 x 10 meters)

Binary image

Rectangular

decomposition

Figure 2: Examples of ﬂoor plans of real environments from the Gibson dataset (Xia et al., 2018). After extracting the

binary images of the navigable areas (ﬁrst row), each ﬂoor plan was decomposed into a set of rectangles (second row). This

representation simpliﬁes the process of importing ﬂoor plans into the proposed simulator. The ﬂoor plans have been scaled

for visualization purposes, their actual size is indicated in the text above each image.

2021), which employed unrealistic human dynamics,

this work extends the simulator introduced in (Chen

et al., 2019) to support advanced human simulations

for recreating complex navigation environments. The

simulator allows dynamic environment generation or

the import of real ﬂoor plans and incorporates a lidar

sensor for obstacle perception. Similar to the “person

sensor”, the signal from the lidar is less affected by

the potential domain gap between the simulated and

real world. Navigation models trained under these

conditions are better equipped to operate in realistic

settings, facilitating the learning of robust and sophis-

ticated navigation policies.

3 TRAIN NAVIGATION MODELS

IN COMPLEX

ENVIRONMENTS

In this section, we provide a detailed discussion on

the proposed simulator designed for training naviga-

tion policies in environments that include both static

components (such as the environment’s layout) and

dynamic components (such as moving humans). We

will release the code of the proposed simulator, along

with the dataset of 2D ﬂoor plans from real-world en-

vironments.

3.1 The CrowdSim++ Simulator

The proposed CrowdSim++ is an enhanced version of

the open-source simulator introduced in (Chen et al.,

2019). This simulator represents the environment on a

2D plane, viewed from a top-down perspective. Fig-

ure 1 displays two examples of navigation episodes

performed within the simulator. Both the robot and

humans are depicted as circles, moving freely on the

2D plane and guided toward their destinations by

the respective navigation policies. The simulator na-

tively implements the Gym library, enabling the use

of numerous Gym-compliant policy optimization al-

gorithms available in the state of the art. The layouts

of the environments used in simulation were recre-

ated using rectangles of varying dimensions. When

combined, these rectangles form more complex ge-

ometries. Examples of the generated environments

are depicted in Figure 2. The proposed system allows

for the creation of two distinct types of scenarios:

1) randomly generated environments with a variable

number of obstacles; 2) environments recreated based

on the ﬂoor plans of real locations. The construction

methods for these scenarios are described below.

Randomly Generated Environments. This strat-

egy offers a practical solution for constructing syn-

thetic environments with varying geometry and com-

plexity, crucial for training robust navigation poli-

cies capable of generalizing to novel environments.

Example of such environments are showed in Fig-

ure 1. The selection of obstacle count and size al-

lows for tailoring the complexity level encountered by

the navigation model during both training and test-

ing phases. During environment generation, obsta-

cles are strategically placed to maintain a margin from

the environment boundaries and each other. This ap-

proach ensures that invalid conﬁgurations, such as

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

536

edge-concentrated (simple layout) or obstructed en-

vironments lacking sufﬁcient navigation space (parti-

tioned environment), are avoided. This study explores

various types of randomly generated environments,

progressively increasing in complexity.

Real Environments. The use of ﬂoor plans de-

rived from real environments holds signiﬁcant impor-

tance as it allows navigation models to learn poli-

cies directly from observations of complex geome-

tries resembling potential deployment environments

for robots. Realistic ﬂoor plans also prove beneﬁcial

in scenarios requiring specialized navigation models

tailored to speciﬁc environments or when early per-

formance estimates are necessary before real-world

testing. Figure 2 displays examples of the ﬂoor plans,

represented as binary images, along with their respec-

tive decomposition into rectangles. In this study, we

focused on the 3D indoor environments from the Gib-

son dataset (Xia et al., 2018), comprising 572 real

buildings reconstructed using advanced 3D scanning

techniques. Given that our simulator supports 2D

ﬂoor plans, a preprocessing step was required to adapt

these environments. To achieve this, we imported the

3D models into the Habitat simulator (Savva et al.,

2019) and developed a custom script to extract the

navigation meshes (navmeshes), i.e. polygon meshes

deﬁning navigable areas for agents. Optimal results

were achieved by setting the virtual agent dimensions

to a height of 0.1 meters and a diameter of 0.2 meters.

For multi-level buildings, only ground ﬂoor rooms

were included. Navmesh data was subsequently con-

verted into binary images to differentiate navigable

areas (white) from obstacles (black), with each pixel

representing 0.1 square meters. To enhance navigabil-

ity while preserving geometry, morphological opera-

tors were applied, notably two iterations of the clos-

ing operation, yielding satisfactory results. Finally, to

import the ﬂoor plan into the simulator, we employed

a rectangular decomposition algorithm

. This trans-

formed non-navigable areas of the 2D binary image

into a set of variable-sized rectangles. The algorithm

minimizes the number of rectangles while maximiz-

ing coverage of the speciﬁc area. Following gener-

ation, the rectangles were sorted based on area size

(from largest to smallest). We then set a threshold of

30 as the maximum number of rectangles per ﬂoor

plan for use during navigation simulations. In our

study, we adopted the “medium” data split of the Gib-

son dataset as proposed in (Xia et al., 2018), which

contains a subset of environments of enhanced qual-

ity. Throughout the transformation process of these

ﬂoor plans, some were excluded due to lack of navi-

We used the Adaptive-Boxes decomposition algorithm

https://github.com/jnfran92/adaptive-boxes

(𝑝

, 𝑝

, 𝑣

, 𝑟)

2D lidar sensor (45 beams, 5m, 180d) and “person sensor”

Figure 3: Visualization of the sensors used by the robot.

The purple segments originating from the robot (yellow cir-

cle) represent the laser beams of the lidar sensor. The red

dots on the obstacles indicate the distances perceived by

each laser beam. In green, the ”person sensor” is high-

lighted, which provides information on the position, speed,

and space occupied by the humans.

gability or disconnected areas. The resulting dataset

is a curated collection comprising 120 distinct envi-

ronments.

Robot’s Perception. Effective navigation for a robot

relies on having the appropriate sensors to perceive

its environment accurately. In this regard, we ex-

tended the CrowdSim simulator (Chen et al., 2019)

to support a 2D lidar sensor, which allows the robot

to get detailed and real-time information about obsta-

cles, walls, and other structural elements of the envi-

ronment. Figure 3 provides a graphical representation

of the lidar sensor. Speciﬁcally, it measures the dis-

tance to objects by calculating the time it takes for

the laser pulses to reﬂect back from surfaces. The

implemented sensor can be customized by adjusting

parameters such as the number of laser beams, maxi-

mum perceived distance, and ﬁeld of view (up to 360

degrees). An obstacle, whether static or dynamic, is

detected by a laser beam if there is an intersection

between the beam and the obstacle. In such cases,

the measured distance is that of the closest intersec-

tion point to the robot. We employed the Shapely

library for efﬁcient management of geometric inter-

sections. Additionally, we implemented a strategy to

enhance measurement efﬁciency by excluding obsta-

cles beyond the sensor’s maximum range.

We used the Shapely library https://github.com/

shapely/shapely

CrowdSim++: Unifying Crowd Navigation and Obstacle Avoidance

537

4 EXPERIMENTAL SETTINGS

AND RESULTS

In this section, we discuss the characteristics of the

environments used for training navigation policies,

the considered navigation models, and the details of

the training and evaluation setup.

4.1 Environments and Simulation

To test our hypothesis that training navigation policies

in complex, crowded environments leads to superior

performance, we designed a series of navigation envi-

ronments with varying levels of difﬁculty.

Generated Environments. For randomly generated

environments, we trained and tested the navigation

models in scenarios with one and three obstacles. We

set each environment to a ﬁxed size of 10 ×10 meters.

Within this space, we generated rectangular obstacles

with variable dimensions, with side lengths randomly

chosen in the range [3,6] meters. In all cases, we

added four additional obstacles along the edges of the

environment. This served two purposes: 1) prevent-

ing agents from navigating outside the environment

without triggering a collision event; 2) enabling the

lidar sensor to accurately detect the boundaries of the

navigable space. Subsequently, To ensure challenging

navigation tasks, we placed agents’ starting positions

and goals at least 7 meters apart (70% of the envi-

ronment’s longest side). To enhance simulation efﬁ-

ciency, we pre-computed obstacle layouts and agent

start/goal positions, loading them as needed instead

of generating them in real-time.

For each scenario, we created 100 unique environ-

ments for both the training and test sets. Each en-

vironment supports thousands of possible navigation

episodes. This setup allows the navigation model to

learn from a diverse range of similarly complex envi-

ronments, enhancing its ability to generalize to new

settings. The number of episodes per environment is

dynamically calculated at the start of each training or

testing session, based on the total number of naviga-

tion episodes.

Real Environments. Real ﬂoor plans provide an op-

portunity to develop navigation policies for highly

complex environments. In our setup, we focused on

training specialized models for navigation in speciﬁc

environments where ﬂoor plans are available in ad-

vance. This scenario is common in practice, as robots

often need in-depth knowledge of environmental lay-

outs to operate with high accuracy. Moreover, this ap-

proach aligns with classic navigation systems that re-

quire access to environmental maps (Durrant-Whyte

and Bailey, 2006). Speciﬁcally, we selected three real

environments from our proposed set of ﬂoor plans,

representing easy, medium, and hard scenarios with

increasing levels of difﬁculty. Figure 2 illustrates

these ﬂoor plans. For each environment, we sam-

pled two disjoint sets of trajectories: one for training

and one for testing. As with the random obstacle sce-

narios, we pre-computed these trajectories to enhance

simulation efﬁciency.

Simulation Details. Our simulations feature environ-

ments with varying numbers of human agents dur-

ing the training phase. For evaluation purposes, we

consistently use scenes containing ﬁve human agents.

All agents, including the robot, are represented as

circular entities with a radius of 0.3 meters. Hu-

mans are guided by the ORCA policy (Van den Berg

et al., 2008). We assume that the robot has holo-

nomic kinematics and that it is invisible to humans.

Navigation is formulated as a point-goal navigation

task (M

oller et al., 2021). At each step, the robot re-

ceives a new observation from the environment, up-

dates the information on the goal to be reached and

chooses one of 16 discrete actions. The observation

includes the robot’s self state w

and, depending on

the type of input considered (as we will see in Sec-

tion 4.2), the lidar observation l

, the states of humans

= [u

,..., u

] assuming a total number of n hu-

mans, or both of them. All data are expressed in the

local reference frame of the robot. The robot state w

consists of the robot’s distance to goal d

, preferred

speed v

pre f

, velocity (v

), radius r. Each human

state u

consists of the human’s position



, p



, ve-

locity





and radius r

. The lidar observation l

consists of n

range measurements indicating the dis-

tance to the closest obstacle from each beam’s direc-

tion. The action space consists of 2 linear velocities

exponentially spaced between (0, v

pre f

]m/s and 8 di-

rections evenly spaced between [0,2π) degrees. The

navigation episode ends if the robot reaches the goal

within a radius of 0.2 meters, when a collision occurs

or when the maximum simulation time is reached.

Each step is equivalent to an execution time of 0.25s

and the maximum simulation time for an episode is

25s.

4.2 Navigation Models

In our study, we considered two popular Reinforce-

ment Learning-based models for crowd navigation:

CADRL (Chen et al., 2017) and SARL (Chen et al.,

2019). In their original formulation, both models rely

on direct input from a “person sensor” that provides

information on the position, velocity, and physical di-

mensions of nearby humans. SARL represents an

improvement over CADRL as it adopts a more ad-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

538

vanced input data aggregation system based on a self-

attention system, to better capture the robot’s interac-

tions with the most relevant humans. In this work, we

modiﬁed the CADRL and SARL models to process

2D lidar signals, allowing us to evaluate their per-

formance in complex, crowded environments under

various sensor conﬁgurations. We conducted exper-

iments to assess the navigation models’ capabilities

when equipped with: 1) “person sensor” only; 2) lidar

sensor only; 3) a combination of “person sensor” and

lidar sensor. From now on, we will refer to these nav-

igation models as CADRL++ and SARL++, respec-

tively. For the lidar conﬁguration, we optimized sen-

sor parameters to balance environmental perception

and simulation efﬁciency. We set the Field of View

(FOV) to 180 degrees, covering the robot’s frontal

semicircle. The lidar uses 45 laser beams and has a

maximum perceptible distance of 5 meters.

4.3 Training and Evaluation Setup

Following the setup in (Chen et al., 2019), the nav-

igation policies were bootstrapped from the demon-

strations provided by ORCA (Van den Berg et al.,

2008) using imitation learning, for a total of 3k navi-

gation episodes. Subsequently, they were trained us-

ing Reinforcement Learning for 10k episodes, using

the ε−greedy exploration strategy, with ε decaying

linearly from 0.9 to 0.5 over the ﬁrst 5k navigation

episodes. The discount factor γ and the learning rate

were set to 0.9 and 0.001, respectively. We found the

reward function deﬁned in (Chen et al., 2019) suit-

able to award task accomplishments while penaliz-

ing collisions or uncomfortable distances. The pol-

icy is trained by the temporal-difference method with

standard experience replay and ﬁxed target network

techniques (Chen et al., 2019; Chen et al., 2017).

Please refer to Section 1 of the supplementary mate-

rial for a more detailed formulation of the Reinforce-

ment Learning framework adopted.

In our evaluation process, we assessed naviga-

tion performance across 500 trajectories for each nav-

igation scenario. These trajectories were sampled

from test environments obtained as indicated in Sec-

tion 4.1. Speciﬁcally, for the randomly generated

environments, we conducted tests under two distinct

scenarios. The ﬁrst scenario involved environments

containing a single obstacle, while the second one fea-

tured environments with three obstacles. In contrast,

our evaluation of real ﬂoor plans utilized the same en-

vironmental layout, but we employed a separate set

of navigation trajectories. All scenarios included ﬁve

human agents.

Our evaluation employed several key metrics to

provide a comprehensive evaluation of the naviga-

tion system’s performance, safety, efﬁciency, and so-

cial awareness: 1) success rate quantiﬁes the aver-

age proportion of episodes completed without colli-

sions. It provides insight into the overall effectiveness

of the navigation system in safely reaching its destina-

tion; 2) collision rate measures the average fraction of

episodes that concluded with the robot colliding with

either humans or obstacles. This metric helps iden-

tify the frequency of navigation failures due to phys-

ical interactions; 3) timeout rate represents the aver-

age proportion of episodes that reached the maximum

simulation time without completion. It indicates sce-

narios where the robot was unable to reach its goal

within the speciﬁed time constraints; 4) navigation

time calculates the average time taken by successful

episodes to reach the goal. This metric allows for a

more precise evaluation of the policy’s performance

when it operates as intended; 5) danger frequency cal-

culates the average fraction of navigation steps during

which the robot came within 0.2 meters (referred to as

the discomfort distance) of a human. This metric as-

sesses the robot’s ability to maintain comfortable dis-

tances from humans during navigation.

5 EXPERIMENTS

In the following experiments, we aim to demonstrate

that existing navigation models can beneﬁt from train-

ing in environments with both complex layouts and

moving humans, particularly when conﬁgured to re-

ceive input signals from additional sensors.

Results for Generated Environments. Table 1

presents the results of navigation models tested in en-

vironments featuring one obstacle and ﬁve humans.

Each line indicates the type of navigation model,

the used sensors (consistent across both the train-

ing and testing phases), the number of obstacles, and

whether humans were present during training (we

ﬁxed the number of humans to ﬁve). As expected,

the CADRL++ navigation model shows limited per-

formance when trained in obstacle-free environments

(lines 1-3). For the model equipped only with a “per-

son sensor”, the information on human behavior was

insufﬁcient to infer the presence of obstacles, which

remain invisible. The model using only the lidar sen-

sor (line 2) was able to perceive static obstacles and

reduce the collision rate compared to the model us-

ing only the “person sensor” (line 2 vs. 1, 0.3240 vs.

0.4788 collision rate). However, it failed to accurately

detect and avoid humans, as indicated by the higher

recorded danger frequency. Moreover, its success rate

CrowdSim++: Unifying Crowd Navigation and Obstacle Avoidance

539

Table 1: Performance of the considered navigation models, tested in environments with one obstacle and ﬁve humans. The

policies were trained in scenarios with a variable number of obstacles (ranging from 0 to 3), with or without humans, and

using the “person sensor,” the lidar sensor, or both. The results conﬁrm that training in crowded environments with a greater

number of obstacles leads to better outcomes. Additionally, using both sensors is essential for optimal performance.

Test environments: 1 obstacle, 5 humans

model sensor(s)

# of

obstacles humans

success

rate (↑)

collision

rate (↓)

timeout

rate (↓)

nav.

time (↓)

danger

freq. (↓)

1 CADRL person 0 ✓ 0.4550 0.4788 0.0662 10.2282 0.0766

2 CADRL++ lidar 0 ✓ 0.2560 0.3240 0.4200 17.5719 0.1232

3 CADRL++ lidar+person 0 ✓ 0.4932 0.2500 0.2568 13.5847 0.0946

4 CADRL++ lidar 1 ✗ 0.4824 0.1480 0.3696 16.1093 0.0871

5 CADRL++ lidar 3 ✗ 0.5644 0.1900 0.2456 12.3953 0.1664

6 CADRL++ person 1 ✓ 0.3124 0.4800 0.2076 12.8515 0.0670

7 CADRL++ person 3 ✓ 0.0120 0.1720 0.8160 25.0000 0.0471

8 CADRL++ lidar 1 ✓ 0.5134 0.2176 0.2690 16.0571 0.1104

9 CADRL++ lidar 3 ✓ 0.6027 0.3024 0.0949 12.6989 0.1708

10 CADRL++ lidar+person 1 ✓ 0.6741 0.1960 0.1299 12.9387 0.0896

11 CADRL++ lidar+person 3 ✓ 0.7172 0.1510 0.1318 11.2958 0.1659

12 SARL person 0 ✓ 0.4727 0.4312 0.0961 9.7005 0.0183

13 SARL++ lidar+person 0 ✓ 0.5688 0.2880 0.1432 10.9796 0.0150

14 SARL++ person 1 ✓ 0.3514 0.3390 0.3096 12.2493 0.0252

15 SARL++ person 3 ✓ 0.0520 0.0900 0.8580 21.8550 0.0624

16 SARL++ lidar+person 1 ✓ 0.6934 0.2295 0.0771 10.9218 0.0272

17 SARL++ lidar+person 3 ✓ 0.7498 0.1560 0.0942 11.5688 0.0651

remains limited because it was not exposed to envi-

ronments with obstacles during training. Adding a

“person sensor” to the existing lidar sensor signiﬁ-

cantly improved navigation performance (line 3). By

explicitly detecting the position and speed of humans,

the model was able to reduce both the collision rate

and the frequency of invading human comfort zones.

Training the models equipped with a lidar sensor in

environments with obstacles but without humans (line

4-5) further reduced the collision rate, even if humans

are not seen during training, suggesting that the li-

dar can compensate for the lack of a dedicated “per-

son sensor”. Using a more complex environment with

three obstacles led to a higher success rate (line 5).

For navigation models trained in the presence of

humans and with a variable number of obstacles (lines

6-11), we observe that more complex environments

are beneﬁcial when the robot uses appropriate sen-

sors. Speciﬁcally, when only the “person sensor” is

used, performance is very limited (line 7, CADRL++

with three obstacles). However, when the lidar sen-

sor is used (lines 8-9), performance improves drasti-

cally. In the scenario with three obstacles, the navi-

gation model learned to navigate the complex geome-

tries, although it showed some difﬁculty in managing

human dynamics, as indicated by a higher danger fre-

quency. The best-performing CADRL++ navigation

models are those equipped with both “person” and li-

dar sensors (lines 10-11). Among these, the model

trained in more complex environments achieved the

highest success rate (0.7172). The SARL++ naviga-

tion model conﬁrmed its superior ability in capturing

human-robot relationships.

When analyzing the performance of SARL++

(lines 12-17), it consistently outperformed

CADRL++ across all training scenarios. It achieved

the best success rate (0.7498) in the scenario where

both the “person sensor” and lidar sensor were used,

and the training took place in environments with

three obstacles and the presence of humans.

Navigation models tested in environments with

three obstacles and ﬁve humans achieved perfor-

mances that mirror the trends observed in Table 1 but

they are generally inferior. Please refer to Section 2

of the supplementary material for a more detailed dis-

cussion.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

540

Navigation episodes in real-world floor plans

“Easy” environment

“Medium” environment

“Hard” environment

Humans

trajectories

Human

Robot

without lidar

Robots

trajectories

Robot

with lidar

Obstacle

Humans

trajectories

Human

Robot

without lidar

Robots

trajectories

Robot

with lidar

Obstacle

Figure 4: Examples of navigation episodes in real ﬂoor plans using the SARL++ model in the easy, medium, and hard

scenarios. The robot equipped with both “person” and lidar sensors (blue circle, red trajectory) successfully navigates through

narrow and crowded areas. In contrast, using the same navigation model without the lidar sensor, the robot collides with

the ﬁrst obstacle encountered (yellow circle, gray trajectory). The ﬁgure represents snapshots of the concluded navigation

episodes.

Overall, the results conﬁrm that navigation models

can learn advanced navigation policies when trained

in complex environments. Training in crowded en-

vironments with a greater number of obstacles leads

to better results and more robust models. Addition-

ally, the use of specialized sensors is advantageous

for capturing the distinct properties of static and dy-

namic obstacles.

Results for Real Environments. Table 2 reports the

results obtained by SARL++ with both “person sen-

sor” and lidar sensor, trained and tested on three real

ﬂoor plans classiﬁed as easy, medium, and hard. The

ﬂoor plans are showed in Figure 4. This navigation

model achieved the best results in the previous eval-

uations. In the easy scenario, SARL++ performed

well, with almost no navigation episodes ending in

timeout, indicating that a robust navigation policy had

been learned, capable of guiding the robot to its desti-

nation efﬁciently. However, the success rate dropped

in the medium and hard scenarios, with the hard sce-

nario showing the highest timeout rate. Similar to the

test conducted in environments with three obstacles,

this performance decline may be due to the use of a

less efﬁcient RL algorithm, which requires a signif-

icantly higher number of training episodes to handle

more challenging environments effectively.

Nonetheless, the average results conﬁrm that navi-

gation policies beneﬁt from exposure to complex real

ﬂoor plans and that training specialized models for

speciﬁc ﬂoor plans is an effective strategy.

Table 2: Navigation performance in real-world ﬂoor plans.

The evaluation focused on the SARL++ model, equipped

with both a person sensor and lidar sensor. Refer to the text

for details.

model sensor(s)

real ﬂoor

plan

humans

success

rate

collision

rate

timeout

rate

SARL++ lidar+person easy ✓ 0.8436 0.1433 0.0131

SARL++ lidar+person medium ✓ 0.7678 0.1842 0.0480

SARL++ lidar+person hard ✓ 0.6610 0.1079 0.2311

avg. 0.7575 0.1451 0.0974

6 CONCLUSION

In this study, we introduced CrowdSim++, an ad-

vanced simulator designed to train navigation mod-

els under realistic conditions. Through experimenta-

tion and comparative analysis, we conﬁrmed the im-

portance of training in complex scenarios involving

humans to develop robust navigation policies suitable

for real-world deployment. We anticipate that our

proposed simulator and dataset of realistic ﬂoor plans

will catalyze further advancements in autonomous

navigation systems.

ACKNOWLEDGEMENTS

This research has been supported by the project Fu-

ture Artiﬁcial Intelligence Research (FAIR) – PNRR

MUR Cod. PE0000013 - CUP: E63C22001940006

CrowdSim++: Unifying Crowd Navigation and Obstacle Avoidance

541

REFERENCES

Bojarski, M., Del Testa, D., Dworakowski, D., Firner,

B., Flepp, B., Goyal, P., Jackel, L. D., Monfort,

M., Muller, U., Zhang, J., et al. (2016). End to

end learning for self-driving cars. arXiv preprint

arXiv:1604.07316.

Cancelli, E., Campari, T., Seraﬁni, L., Chang, A. X., and

Ballan, L. (2023). Exploiting proximity-aware tasks

for embodied social navigation. In International Con-

ference on Computer Vision (ICCV). IEEE.

Chen, B., Sax, A., Lewis, F., Armeni, I., Savarese, S.,

Zamir, A., Malik, J., and Pinto, L. (2021). Robust

policies via mid-level visual representations: An ex-

perimental study in manipulation and navigation. In

Kober, J., Ramos, F., and Tomlin, C., editors, Pro-

ceedings of the 2020 Conference on Robot Learning,

volume 155, pages 2328–2346.

Chen, C., Liu, Y., Kreiss, S., and Alahi, A. (2019). Crowd-

robot interaction: Crowd-aware robot navigation with

attention-based deep reinforcement learning. In In-

ternational Conference on Robotics and Automation

(ICRA).

Chen, T. L., Ciocarlie, M., Cousins, S., Grice, P. M.,

Hawkins, K., Hsiao, K., Kemp, C. C., King, C.-H.,

Lazewatsky, D. A., Leeper, A. E., et al. (2013). Robots

for humanity: using assistive robotics to empower

people with disabilities. IEEE Robotics & Automation

Magazine, 20(1).

Chen, Y. F., Liu, M., Everett, M., and How, J. P. (2017). De-

centralized non-communicating multiagent collision

avoidance with deep reinforcement learning. In In-

ternational Conference on Robotics and Automation

(ICRA).

Dong, M. and Zhang, J. (2023). A review of robotic grasp

detection technology. Robotica.

Durrant-Whyte, H. and Bailey, T. (2006). Simultaneous lo-

calization and mapping: part i. IEEE robotics & au-

tomation magazine.

Khanna, M., Mao, Y., Jiang, H., Haresh, S., Shacklett, B.,

Batra, D., Clegg, A., Undersander, E., Chang, A. X.,

and Savva, M. (2024). Habitat synthetic scenes dataset

(hssd-200): An analysis of 3d scene scale and realism

tradeoffs for objectgoal navigation. In Computer Vi-

sion and Pattern Recognition (CVPR).

Kuderer, M., Kretzschmar, H., Sprunk, C., and Burgard, W.

(2012). Feature-based prediction of trajectories for so-

cially compliant navigation. In Robotics: science and

systems, volume 8.

Liu, L., Dugas, D., Cesari, G., Siegwart, R., and Dub

e, R.

(2020). Robot navigation in crowded environments

using deep reinforcement learning. In 2020 IEEE/RSJ

International Conference on Intelligent Robots and

Systems (IROS).

Mavrogiannis, C., Baldini, F., Wang, A., Zhao, D., Traut-

man, P., Steinfeld, A., and Oh, J. (2023). Core chal-

lenges of social robot navigation: A survey. ACM

Transactions on Human-Robot Interaction, 12(3).

Miller, D. P. (2006). Assistive robotics: an overview. As-

sistive Technology and Artiﬁcial Intelligence: Appli-

cations in Robotics, User Interfaces and Natural Lan-

guage Processing.

oller, R., Furnari, A., Battiato, S., H

arm

a, A., and

Farinella, G. M. (2021). A survey on human-aware

robot navigation. Robotics and Autonomous Systems,

145.

Otte, M. W. (2015). A survey of machine learning ap-

proaches to robotic path-planning. University of Col-

orado at Boulder, Boulder.

erez-D’Arpino, C., Liu, C., Goebel, P., Mart

ın-Mart

ın, R.,

and Savarese, S. (2021). Robot navigation in con-

strained pedestrian environments using reinforcement

learning. In International Conference on Robotics and

Automation (ICRA).

Quigley, M., Conley, K., Gerkey, B., Faust, J., Foote, T.,

Leibs, J., Wheeler, R., Ng, A. Y., et al. (2009). Ros: an

open-source robot operating system. In ICRA work-

shop on open source software, volume 3.

Rosano, M., Furnari, A., Gulino, L., and Farinella, G. M.

(2021). On embodied visual navigation in real envi-

ronments through habitat. In International Conference

on Pattern Recognition (ICPR).

Rosano, M., Furnari, A., Gulino, L., Santoro, C., and

Farinella, G. M. (2023). Image-based navigation in

real-world environments via multiple mid-level repre-

sentations: Fusion models, benchmark and efﬁcient

evaluation. Autonomous Robots, 47(8).

Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans,

E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J.,

et al. (2019). Habitat: A platform for embodied ai

research. In International Conference on Computer

Vision (ICCV).

Staranowicz, A. and Mariottini, G. L. (2011). A survey and

comparison of commercial and open-source robotic

simulator software. In Proceedings of the 4th Interna-

tional Conference on PErvasive Technologies Related

to Assistive Environments.

Van den Berg, J., Lin, M., and Manocha, D. (2008). Recip-

rocal velocity obstacles for real-time multi-agent nav-

igation. In International Conference on Robotics and

Automation (ICRA).

Xia, F., Zamir, A. R., He, Z., Sax, A., Malik, J., and

Savarese, S. (2018). Gibson env: Real-world percep-

tion for embodied agents. In Computer Vision and

Pattern Recognition (CVPR).

Yaar, A., Furnari, A., Rosano, M., H

arm

a, A., and Farinella,

G. M. (2024). Finding and navigating to humans in

complex environments for assistive task. In Inter-

national Conference on Computer Vision Theory and

Applications (VISAPP).

Zhu, K. and Zhang, T. (2021). Deep reinforcement learning

based mobile robot navigation: A review. Tsinghua

Science and Technology, 26(5).

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

542