Navigation of a two-wheel differential drive robot

in a partially unknown environment

Breno P. de Meneses ∗ Gabriel H. V. da Silva ∗ Lara R. Sobral ∗
Mateus S. Marques ∗ Rodrigo T. de Araujo ∗ Antonio M.N. Lima ∗∗
∗
DEE-UFCG
∗∗
Department of Electrical Engineering (DEE), Universidade Federal
de Campina Grande (UFCG), PB, Brazil
([email protected])

Abstract: This paper proposes the use of a behavior-based control strategy for collision-free
navigation of a two-wheel differential drive mobile robot in partially unknown environments. The
proposed strategy is hierarchical, and the design of the high-level control layer (behavioral space)
is based on Mamdani fuzzy logic inference, while the design of the low-level control layer (joint
space) is based on Lyapunov stability theory. The collision and deadlock avoidance technique is
based on subgoals that are evaluated in the behavioral space and supplied as references to the
joint space controller. The in silico test study was executed by using the CoppeliaSim which is
a simulation framework used for the prototyping, development and verification of robot systems
and algorithms that is widely used by the robotics community. Several (robot initial and final
poses, walls, corridors, make-like regions, and loose objects placement) test scenarios, and a
comparative study with previous and related work were performed. The test results show the
proposed strategy provided a smooth and shorter path in all cases. This improvement is basically
related with the combined use of the subgoal technique and a Lyapunov based controller. In
summary, the results corroborate the correctness of the methodology adopted in the design of
collision-free navigation for differential drive mobile robots.

Keywords: Behavior-Based Systems, Sensor-based Control, Control Architectures,

Programming, Fuzzy Control and Nonlinear Systems

1. INTRODUCTION In general, when an autonomous robot must navigate

in a relatively complex and unknown environment, one
An autonomous mobile robot is a programmable device prefers navigation strategies based on reactive control
that uses proprioceptive and exteriorceptive sensor mea- concepts (De Silva and Ekanayake, 2008). The combina-
surements to gather data from the surrounding environ- tion of the different behaviors (behavior-based control)
ment, and process it to plan its motion and execute given that the robot should exhibit in a given environment is
tasks, and at the same time avoid collisions with objects a relatively simple and intuitive reactive control strat-
that share the same space. Whenever a detailed map of egy. In the behavior-based control, one chooses a set of
the environment is known before hand, one may plan the robot behaviors, e.g., “goal seeking”, “obstacle avoidance”,
robot’s navigation to avoid collisions with static objects. “deadlock disarming”, and a technique to combine them.
Besides, the map-based navigation, rely on an accurate The selection of a given behavior is evaluated based on
and up-to-date representation of the environment, which the measurements provided by the onboard robot sensors.
may not be available in unknown or partially unknown Fuzzy Logic (Bao et al., 2009; Van Nguyen et al., 2017a,b),
environments. Arbitration (Brooks, 1986; Baumann et al., 2022), Blend-
ing (Adriansyah, 2014; Ramakrishna Pandian et al., 2021),
If unpredictable or dynamic objects appear in the robot’s Potential Fields (Khatib, 1986; Kim et al., 2016), Behavior
path, the navigation planner must include a collision avoid- Trees (Colledanchise and Ögren, 2018), Finite State Ma-
ance strategy based on the on-board robot sensing capa- chines (Petrovic, 2008; Bozzi et al., 2022), Reinforcement
bilities. This is true in applications like, for instance, au- Learning (Cherroun and Boumehraz, 2012; Sutton and
tonomous vehicles, warehouse automation, robot vacuum Barto, 2018) and Evolutionary Algorithms (Shen, 2013;
cleaners, to name just a few. When a mobile robot navi- Sathiya and Chinnadurai, 2019) have been used individ-
gates autonomously in any given area, at least a collision ually or in combination to design behavior-based control
and deadlock avoidance technique is required to ensure systems for mobile robots, allowing for flexible and adap-
safe and efficient functioning. tive robot behaviors in various environments and tasks.
⋆ This work was carried out with the support of the Department Among the mentioned methods, fuzzy logic was chosen
of Electrical Engineering, Center for Electrical Engineering and In- because it provides an effective technique to deal with
formatics (CEEI), Federal University of Campina Grande, Campina uncertainty by means of flexible linguistic categories and
Grande, PB.
logical rules of inference. This reduces the difference be-
tween human reasoning and numerical application, making
it possible to define control systems through linguistic
rules, which are adapted according to behavior, as in
this context. In addition to its effectiveness as a control
technique, fuzzy logic requires less formal knowledge of
the plant and the environment in which it will be applied,
in comparison to the so-called standard control system
approach, allowing specific movements to be carried out
with a simplified implementation.
Fuzzy behavior-based architectures for two-wheel differen-
tial drive robot navigation in partially unknown environ-
ments with static objects are proposed in (Bao et al., 2009)
and (Van Nguyen et al., 2017a,b). The so called subgoal
technique has been proposed in (Ye and Webb, 2009) for
performance optimization. In general, the formulation of
a behavior-based control problem does not deal with the
low-level control, i.e., robot’s joints control. Without the
robot’s low-level control layer there is no guarantee that
the actual motion of the robot will be exactly as smooth Figure 1. Schematic of the position and orientation of the
as expected by the behavioral control layer (Panahandeh vehicle with respect to the final goal.
et al., 2019). The use of Lyapunov-based low-level con-
troller for moving target tracking with a fuzzy strategy for Besides the basic assumptions regarding the environment,
obstacle avoidance has been proposed in (Benbouabdallah we also consider that
and Zhu, 2013; Kubo et al., 2020). However, the collision
avoidance strategy is restricted to point objects, and thus • For designing both the high level and the low-level
cannot deal with deadlocks. controller, one considers that the two-wheel differen-
tial drive robot is represented by the so called unicycle
In this paper, we include a high-level control layer based model, which is a third-order two-input non-linear
on Fuzzy Logic that exploits the subgoal approach to kinematic model (Lynch and Park, 2017); from now
determine robot motion routes. Also in this layer, the fuzzy on we will denote this model as our action model.
controller allows the vehicle to develop different types of • For testing the design solution one considers that
behaviors (e.g., obstacle avoidance, tracking, and deadlock the two-wheel differential drive robot with casters
disarming) while heading to a goal (go to goal). The design (off-centered orientable, not driven wheel joint) is
of the low-level control layer is based on Lyapunov stability represented by a dynamic model (mass, moment of
theory and provides robust stability and convergence. inertia, friction), which is a fourth-order two-input
non-linear model (Dhaouadi and Hatab, 2013); from
2. PROBLEM DESCRIPTION now on we will denote this model as our knowledge
model.
• The testing environment (a virtual world scene in-
Figure 1 shows a schematic illustration of a robot with two-
cluding the floor, walls, corridors, maze-like regions,
wheel differential drive, a final goal, and an environment
obstacles, range sensors, and robot) and the knowl-
with an obstacle, supposedly static and unknown. The
edge model are simulated by using the latest version
objective of this paper is to design a control strategy that
of CoppeliaSim (Rohmer et al., 2013). CoppeliaSim
allows the robot to go to the final goal while safely avoiding
is a simulation framework used for the prototyping,
collision in its path. To do this, the following assumptions
development and verification of robot systems and
are made
algorithms that is widely used by the robotics com-
• The goal (target) and robot pose is known and spec- munity.
ified, meaning that we consider a partially unknown
The proposed control strategy is hierarchical. In the upper
environment.
layer, the behavior-based controller combines the behav-
• The obstacles in the environment are unknown to the
iors and sends them to the lower layer, where a nonlinear
robot, meaning it does not have prior information
controller, tuned by using a genetic algorithm, is used.
about their locations or shapes.
• The robot operates in a two-dimensional space, as
indicated by the schematic illustration in Figure 1.
• The robot is equipped with range sensors that pro- 3. LOW-LEVEL CONTROL LAYER
vide enough information for detecting and localizing
obstacles. 3.1 Lyapunov Controller
• The robot is a non-holonomic system, and no lateral
slip motion is allowed. However, it can move forward, Consider a vehicle positioned at a nonzero distance with
backward, and rotate clockwise and counterclockwise respect to a goal pose, as depicted in Figure 1. As Kubo
around the z-axis (perpendicular to the xy plane et al. (2020) and Ravangard (2015) a Lyapunov controller
depicted in Figure 1). can be used to describe a unicycle-like motion given by
 
ẋ = u cos (ϕ)
ẏ = u sin (ϕ) (1)

ϕ̇ = ω
where x, y, denote the vehicle position, and ϕ its orien-
tation. The control inputs are the linear velocity, u, and
the rotational velocity ω, with respect to the xI × yI
frame. However, for designing the non-linear control law
the vehicle motion will be represented by

ė = −u cos (α)

α̇ = −ω + u sin(α)
e
(2)
θ̇ = u sin(α)

e
where e and α denote the distance and orientation with
respect to the goal, respectively. Now by choosing following
Lyapunov candidate function
1 1
V (e, α) = λe2 + hα2 , λ > 0, h > 0 (3)
2 2
Feedback control laws can be found by minimizing their
time derivatives, causing them to asymptotically converge
to zero and become independent of the values of λ and h.

u = γe cos (α) Figure 2. Simple scene containing only four loose points-
(4)
ω = kα + γ cos (α) sin (α) objects. The red path is associated with Snew and the
These equations guarantees global stability and the bound- blue path is associated with Sold
edness of the state trajectory corresponding to any Table 1. Path length and path travel time
bounded initial condition for κ > 0 and γ > 0 (Aicardi comparison (Figure 2)
et al., 1995).
Path length (m) Path travel time (s)
It is worth to point out that the error variables e and α Sold 13.76 14.39
used to compute u and ω depend on the actual goal being Snew 7.28 14.59
tracked. Besides, one has to use the following equations
ωr = 2u+ωL 3.2 Controller gains tuning

2R (5)
ωl = 2u−ωL
2R
The selection of the low-level controller gains was formu-
to determine the left (ωr ) and right (ωr ) wheel joint lated as an optimization problem, and its solution was
velocities (L and R are defined in Figure 1). determined by using a genetic algorithm (Burjorjee, 2007).
The use of a nonlinear low-level controller has been pro- The two gains were coded into a 40-bits word, being 20-
posed by Benbouabdallah and Zhu (2013). However, their bits for κ and 20-bits for γ. The chosen test scenario is
control law is given by depicted in Fig. 3, and the fitness function was defined as

(
cos(β)
u = vT cos(α) − Kv eD cos (α)
(6)
ω = −Kw α − D v
sin (α) + vDT sin (β)
where 
α = θ − Φ
β = θT − Φ (7)
eD = Dd − D


Please see (Benbouabdallah and Zhu, 2013) for the mean-

ing of the terms of the control law. To justify our choice Figure 3. Test scenario used in the controller gains tuning
of the control law given in (4) instead of the one given in procedure.
(6), we conducted a preliminary comparison with a simple
scene containing only loose points-objects. Figure 2 shows 1 1
f itness = + , (8)
this simple scene and the robot navigation under both the Φy Φu
[Snew -(4)] and [Sold -(6)] control laws. In Figure 2 the red where
path is associated with Snew and the blue path is associated
with Sold . Table 1 show that the path length under Sold is Φy = |max(y) − yref | and Φu = |max(u) − uref |. (9)
larger than (≈1.9X) the one observed under Snew , while the This choice of this fitness function aims at ensuring that,
path travel times are essentially the same. For the [Sold - during the robot’s movement toward the target pose, u <
(6)] we used θT = θ = 90◦ , α = β, Kv = 2.07, Kw = uref and y < yref is always satisfied. In the test scenario,
1.49, Dd = 0.2 m , and vT = 0.15 m/s. Whereas, for the the initial robot pose is (0, 0, π/2), and the target pose
[Snew -(4)] we used γ = 0.3 and κ = 1. is (0, 1, −π/2). The genetic algorithm was configured to
use the following hyperparameters: population size = 200, Algorithm 1 Grouping sensor readings
number of generations = 40, crossover probability = 1, and
Input : αgoal , [d1 , · · · , d16 ]
the mutation probability per bit = 0.003. The final gains
Output: Grouped distances along the wind rose with
were γ = 0.328 and κ = 1.08. For these gain values, the
north in the direction of movement.
time-to-target is 5 s, ymax = 0.15 m, and umax = 1.2 m/s.
Require t0
4. HIGH-LEVEL CONTROL LAYER 1: if |αg | > π
2 then
2: dSW ← min(d14 , d15 )
The proposed fuzzy behavior-based architecture for mobile 3: dWE ← min(d1 , d16 )
robot navigation is depicted in Figure 4. For implementing 4: dNW ← min(d2 , d3 )
the high-level control layer strategy one assumes that the 5: dNO ← min(d4 , d5 )
on-board robot sensing capability is a belt of 16 ultrasound 6: dNE ← min(d6 , d7 )
range sensors disposed as depicted in Figure 5. However, 7: dES ← min(d8 , d9 )
the same concepts used here can be applied to other types 8: dSE ← min(d10 , d11 )
of range sensors like for instance infrared and LIDAR. 9: dSO ← min(d12 , d13 )
10: u←u
11: else
12: dSW ← min(d6 , d7 )
13: dWE ← min(d8 , d9 )
14: dNW ← min(d10 , d11 )
15: dNO ← min(d12 , d13 )
16: dNE ← min(d14 , d15 )
17: dES ← min(d1 , d16 )
18: dSE ← min(d2 , d3 )
19: dSO ← min(d4 , d5 )
Figure 4. Fuzzy behavior-based architecture for mobile 20: u ← −u
robot navigation. 21: return dSW, dWE, dNW, dNO, dNE, dES, dSE, dSO

4.1 Sensor arrangement 4.2 Basic behaviors

The sensors of the belt are grouped in eight groups,
The behavior determination is based on readings of the dis-
according the wind rose orientations, i.e., west, northwest,
tance sensor groups, differing only in the quantity, number
north, northeast, east, southeast, south and southwest, as
of groups and position of the sensors (Van Nguyen et al.,
shown in the Figure 5.
2017a,b; Kuo et al., 2013). In the proposed fuzzy archi-
tecture, the basic behaviors are “goal seeking”, “obstacle
avoidance”, “tracking” and “deadlock”. The membership
functions used are summarized as shown in Figure 6. The
control signals are the linear (u) and angular (ω) velocities.
By using the low-level Lyapunov control law, it is possible
to change the input variables for distance (e) and target
orientation (α), as if the goal was elsewhere.
The linguistic labels for the distances issued from the
ultrasound range sensor and to the target are near (N),
medium (M) and far (F), as shown in Figure 6 (a), (c)
and (d). The target distance range was calculated so that
it would not saturate linear velocity of the robot. For
orientation, the inspiration of wind roses is also used,
Figure 5. Placement of the ultrasound sensors on the robot where SOp and SOn are the two halves representing the
chassis and assignment of the clusters based on the SO, observable in Figure 6 (b).
wind rose orientations. (a) Standard orientation; (b)
Reorientation of the clusters to redefine the front of The rules that define each behavior of the proposed archi-
the robot based on the target orientation. tecture can be summarized as follows:
Goal-Seeking: It is not necessary to implement the fuzzy
The distances from the robot to an obstacle (di ) along the behavior to achieve the goal in the absence of obstacles.
wind rose orientations are determined by combining the This is because its input, the distance and orientation
readings issued from two neighboring ultrasound sensors. to the goal, can be imposed on the low-level controller
At the time of its initialization, based on the quadrant of without the need to perform inference. In this way, the
the goal, shown in the grouping sensor readings algorithm subgoal pointed by this behavior is oriented in the same
(Algorithm 1), the front of the robot is set, which will direction as the main goal and located within a defined
imply either positive (1st or 2nd quadrant) or negative (3rd radius;
or 4th quadrant) linear velocities. Thus, by this grouping
sensor readings the north is always in the same direction Obstacle Avoidance: The fuzzy behavior designed for ob-
of the motion, as shown in Figure 5. stacle collision avoidance does not use the distance read-
 4.3 Subgoals generator

The subgoal generator yields a position relative to the

robot’s frame as input for the low-level velocity controller,
rather than directly producing linear and angular veloci-
ties as outputs from the fuzzy controller. The inputs are
the polar coordinates error variables [e, α], as explained in
section 3. The subgoal is determined based on the rules
defined for each behavior in the modular fuzzy controller.
The rules that define the subgoal for each of the considered
behaviors were summarized in the Section 4.2.
By means of the subgoal generator, new positions will be
generated as virtual goals for the robot to act according
to that particular condition. This can be seen as the robot
seeking multiple sequential targets that collectively will
form the path leading to the final target, as depicted in
Figure 7.
Figure 6. The linguistic labels and Gaussian membership
functions (a) Goal distance; (b) Goal orientation; (c)
Goal distance output of Deadlock behavior; (d) Wind
rose distances from WE, NW, NO, NE, ES, SE, SO
and SW sensor clusters.

ings of all clusters, as this would increase the number of

rules, only those arranged in the direction of movement of
the robot (WE, NW, NO, NE and ES). Also, to obtain
a possible shortest trajectory, the goal direction is always
considered. It is important to note that the front distance
(NO) needs to be small, so small that this behavior is acti-
vated, and its rules are derived from this initial condition.
Thus, the subgoals defined by this behavior are always
oriented either WE or ES of the robot, depending on the
goal orientation and position of the obstacle, i.e. , during
goal-seeking behavior an L-shaped front and left (right)
side obstacle is identified, the subgoal is positioned a short Figure 7. Virtual goals issued by the subgoal generator
distance in the direction of ES (WE); g1 = (t1 , e1 , α1 ) , · · · , gn−1 = (tn−1 , en−1 , αn−1 ) to-
wards the final target but avoiding collisions.
Tracking: An important behavior to design is tracking,
which generally consists of unilateral and bilateral wall- For combining the behaviors we also used fuzzy logic. The
following behaviors, i.e. corridors. This behavior should be input for this block are the wind rose distances from the
complementary to obstacle avoidance behavior, avoiding robot to an obstacle while the outputs are the weights
strictly lateral collisions and following them. Thus, right- that each behavior will have in determining the subgoal
angle lateral sensor groups (WE and ES) are evaluated, that will be provided to the low-level controller. Although
provided there is no frontal obstacle (NO), and target the fuzzy rules be based on (Bao et al., 2009), in the
orientation so that it does not follow corridors when present solution there is no discontinuity since the behavior
unnecessary. Thus, the subgoals will always be within a transition is gradual and thus avoids speed spikes in the
small distance in the NO, WE or ES direction, when the robot motion. This smoothness can be inferred from the
sub-object is behind the wall or in front of it, to the left graphical representation of the rule surface depicted in
or right, respectively; Figure 8.
Deadlock: Deadlock disarming is important for solving
dead-end corridors and u-shaped obstacles. It is possible
to find this solution without describing a specific behavior,
merely avoiding obstacles and following the trail. How-
ever, for better performance, we specify this behavior.
In this paper, a rotation on its own axis clockwise or
counterclockwise is proposed. In the case where the front
and side sensor groups (NO, WE, and ES) are triggered Figure 8. Rule surface for the fuzzy logic controller used
simultaneously and the distance to the obstacle is small, to combine behaviors.
a subgoal is defined such that the motion is towards SO,
and this eventually leads the robot to make a rotation With {ei , αi , Wi }, i ∈ {g, r, o, d} the weighted sums are
clockwise or counterclockwise. computed by
 2006). The ultrasonic sensors are arranged around the
X
e (tk ) = Wi (tk−1 ) ei (tk−1 )
i
robot chassis to provide a 360◦ viewing angle, as shown
X in Figure 5. All sensors in the belt are considered to have
α (tk ) = Wi (tk−1 ) αi (tk−1 ) (10) the same sensing characteristics, i.e., a detection range of
i 3 m and a field of view cone of 15◦ . The gains of the low-
at each sampling instant (tk , k = 1, 2, · · · ) to generate level controller were chosen to be γ = 0.3 and κ = 1, and
the new subgoal along the final goal. The subscripts g, r, were kept constant in all test conditions.
o and d denote the “Goal-Seeking”, “Tracking”, “Obstacle
Avoidance” and “Deadlock Disarming” behaviors, respec- For the sake of comparison, we implemented the work
tively. Each blue circle in the Figure 9 image represents of Bao et al. (2009) (Sold ) and compared it with the
the virtual goal generated by using (10). solution proposed here (Snew ) under the same scenarios
to show the differences with respect to the smoothness
and length of the robot’s path. The three test scenarios
are essentially the same used in Bao et al. (2009) and
Van Nguyen et al. (2017a,b), but we changed the initial
position and orientation of the robot, as well as of the goal.
The reason for choosing these scenarios for demonstration
is due to their different complexities and similarity with
actual scenarios, to cover the different behaviors described
previously.
Figures 10 (a) and (d) are relatively simple environments
without loose obstacles other than the walls. It can be
observed that in both cases Sold and Snew the robot suc-
cessfully navigates, avoiding the obstacles and reaches the
goal. However, when comparing the robot paths (Horiuchi
and Noborio, 2001), it is possible to identify that under
Snew the path length is shorter and smoother than the one
observed under Sold . Table 2 Scene a show that the path
Figure 9. Representation of the position of the virtual goals length and path travel time under Sold are larger than
generated by each of the behaviors (≈2X) as the ones observed under Snew

Table 2. Path length and path travel time

5. IN SILICO TESTS comparison (Figure 10)
To show the effectiveness of the control strategy proposed Path length (m) Path travel time (s)
in this work, a model of the mobile robot, the surrounding Sold 21.58 43.20
Scene a
3D environment and the obstacles were instantiated on Snew 10.81 22.24
a virtual prototyping platform called CoppeliaSim Edu Sold 16.26 44.55
Scene b
V4.5.1 (Rohmer et al., 2013). This open-source platform Snew 13.97 37.45
is widely used by the robotics community and allows the Sold 14.27 27.70
Scene c
Snew 13.34 33.29
creation of scripts for remote communication with the
Sold 15.90 41.80
created scene, controlling individually the elements of the Scene d
Snew 13.89 38.395
scene and allowing the calculation of the robot movements,
simulating the operation of the on-board range sensors and
controlling interactions of the robot with obstacles, floors Although having the same starting point (Figures 10 (a)),
and other relevant elements. The physics engine used was the robot’s orientation under Snew is rotated by 180°,
the Bullet Physics Library (Coumans, 2015), is a robust highlighting the efficacy of using the evaluation of the
and versatile physics simulation library that offers high- quadrant of the objective in the definition of the robot’s
performance collision detection, rigid body dynamics, and front, performing the proper behavior configurations, and
preventing local minimum.
constraint solving capabilities. Finally, we used the Pioneer
3DX (P3DX) from Adept Mobile Robots, which is a two- Figures 10 (b) and (d) show a maze-like situation, .i.e,
motor, two-wheeled differential drive robot with caster a nearly closed environment with restricted space for
wheel (Robotics, 2006). making turns. This type of scenario is built in order to
For the in silico tests it was necessary to create virtual emphasize the smoothness of the robot’s movement. When
worlds (scenes) where the P3DX is inside a room that comparing the trajectories under Sold and under Snew , we
contains several obstacles placed at different positions as see a remarkable difference. In the former, the curves are
shown in Figs. 10(first scene)-(sixth scene). For simplicity sharper, almost scraping the walls, and susceptible to entry
we considered that the instantaneous P3DX pose as ob- into a loop since the walls are very close, resulting in
tained from the ground truth provided by CoppeliaSim is a sequence of behaviors that do not reach the solution.
sent to the control strategy, but on the hand the informa- On the other hand, under Snew the path is smooth, with
tion regarding the obstacles is hid. movements more compatible with obstacle placement. For
the same scenario, a new position for the final goal was
The parameters of the P3DX are R = 0.0975 m, L = assigned so that Snew decides to navigate backwardly. In
0.3810 m, umax = 1.2 m/s, and ωmax = 4.3 rad/s (Robotics, these cases, we can see less effort to align the robot’s pose
Figure 10. Comparison with the robot motion (colored triangles) without (Sold ) and with the low-level control layer
(Snew ) in different scenarios, start point and end goal. The low-level control design is based on Lyapunov stability
theory. (a) Corridor scenario; (b) Maze scenario with two exits; (c) Room scenario with goal in another room; (d)
Maze scenario with one exit.
with the goal pose, and the robot navigates backwardly. it was placed between two less problematic objects in terms
Table 2, Scene b show that the path length and path travel of convergence. 1
time under Sold are larger than (>18%) the ones observed
under Snew 6. CONCLUSION
Figures 10 (c), show environments close to real situations
that rely on room sets and common areas with smaller This paper has presented a behavior-based control strat-
objects compared to walls, such as chairs and centers. In egy for collision-free navigation of a two-wheel differential
these cases, under both Sold and Snew to robot reach the drive mobile robot in partially unknown environments.
objective. Again, under Snew the motion is smoother with The proposed strategy is hierarchical, and the design of the
fewer abrupt curves and far from the walls. As in the other behavioral layer controller is based on fuzzy logic inference,
cases, the goal was allocated in such a position that the while the design of the wheel joint controller is based
robot under Snew navigates backwardly, and furthermore, on Lyapunov stability theory. We have also compared
1 Video demonstration: Robot navigating under Snew in Cop-
peliasim.
two different nonlinear control laws for the wheel joint the alternative following. In Conf. Rec. IEEE/ICRA,
controller. At the behavioral level, the collision and dead- volume 2, 1728–1735.
lock avoidance is achieved by means of selecting proper Khatib, O. (1986). Real-time obstacle avoidance for
subgoals towards the final. The in silico study revealed manipulators and mobile robots. Int. J. Rob. Res., 5(1),
that the proposed strategy yields a mobile robot motion 90–98.
that is smooth, short, and collision and deadlock free. Kim, Y.H., Son, W.S., Park, J.B., and Yoon, T.S. (2016).
Some future works include the validation and refinement Smooth path planning by fusion of artificial potential
of the proposed solution by using a laboratory prototype, field method and collision cone approach. In MATEC
extending the study to consider the dynamic model of Web of Conferences, volume 75, 05004. EDP Sciences.
the obstacle and mobile robot for dealing with energy Kubo, R., Fujii, Y., and Nakamura, H. (2020). Control lya-
constraints, and modifying the low-level control layer gains punov function design for trajectory tracking problems
according to prescribed subgoals. of wheeled mobile robot. IFAC-PapersOnLine, 53(2),
6177–6182.
REFERENCES Kuo, C.H. et al. (2013). Development of a fuzzy logic wall
following controller for steering mobile robots. In Conf.
Adriansyah, A. (2014). Design of context dependent blend- Rec. iFUZZY, 7–12. IEEE.
ing (cdb) in behaviour based robot using particle swarm Lynch, K.M. and Park, F.C. (2017). Modern Robotics: Me-
fuzzy controller (psfc). In 1st International Conference chanics, Planning, and Control. Cambridge University
on Computer Science and Engineering. Sriwijaya Uni- Press.
versity. Panahandeh, P., Alipour, K., Tarvirdizadeh, B., and Hadi,
Aicardi, M., Casalino, G., Bicchi, A., and Balestrino, A. A. (2019). A kinematic lyapunov-based controller to
(1995). Closed loop steering of unicycle like vehicles via posture stabilization of wheeled mobile robots. Mechan-
lyapunov techniques. IEEE Robot. Autom., 2(1), 27–35. ical Systems and Signal Processing, 134, 106319. doi:
Bao, Q.Y., Li, S.M., Shang, W.Y., and An, M.J. (2009). 10.1016/j.ymssp.2019.106319.
A fuzzy behavior-based architecture for mobile robot Petrovic, P. (2008). Evolving behavior coordination for
navigation in unknown environments. In Conf. Rec. mobile robots using distributed finite-state automata.
IEEE/AICI, volume 2, 257–261. INTECH Open Access Publisher.
Baumann, C., Birch, H., and Martinoli, A. (2022). Lever- Ramakrishna Pandian, K.K., Dinh, T.Q., and Marco, J.
aging multi-level modelling to automatically design be- (2021). Modelling and brake blending control for multi-
havioral arbitrators in robotic controllers. In Conf. Rec. drive mode electric two-wheelers. In 24th International
IEEE/IROS, 9318–9325. Conference on Mechatronics Technology (ICMT), 1–6.
Benbouabdallah, K. and Zhu, Q.D. (2013). A behavior- doi:10.1109/ICMT53429.2021.9687134.
based controller for a mobile robot tracking a moving Ravangard, M. (2015). Fuzzy behavior based mobile robot
target in multi-obstacles environment. In Conf. Rec. navigation. In Conf. Rec. 4th CFIS, 1–7. doi:10.1109/
IEEE/IHMSC, volume 2, 418–423. CFIS.2015.7391649.
Bozzi, A., Graffione, S., Sacile, R., and Zero, E. (2022). Robotics, A. (2006). Pioneer 3 operations manual.
Design and implementation of an asynchronous finite Rohmer, E., Singh, S.P., and Freese, M. (2013). V-rep: A
state controller for wheeled mobile robots. In Actuators, versatile and scalable robot simulation framework. In
volume 11, 330. MDPI. doi:10.3390/act11110330. Conf. Rec. IEEE/IROS, 1321–1326.
Brooks, R. (1986). A robust layered control system for a Sathiya, V. and Chinnadurai, M. (2019). Evolutionary
mobile robot. IEEE Robot. Autom., 2(1), 14–23. algorithms-based multi-objective optimal mobile robot
Burjorjee, K. (2007). Speedyga: a fast simple genetic trajectory planning. Robotica, 37(8), 1363–1382.
algorithm. Mathworks Central. Shen, X. (2013). A quantum evolutionary algorithm for
Cherroun, L. and Boumehraz, M. (2012). Intelligent sys- robot path planning in dynamic environment. In Conf.
tems based on reinforcement learning and fuzzy logic Rec. IEEE/CCC, 8061–8065.
approaches, ”application to mobile robotic”. In Inter- Sutton, R.S. and Barto, A.G. (2018). Reinforcement
national Conference on Information Technology and e- learning: An introduction. MIT press.
Services, 1–6. doi:10.1109/ICITeS.2012.6216661. Van Nguyen, T.T., Phung, M.D., and Tran, Q.V. (2017a).
Colledanchise, M. and Ögren, P. (2018). Behavior trees in Behavior-based navigation of mobile robot in unknown
robotics and AI: An introduction. CRC Press. environments using fuzzy logic and multi-objective op-
Coumans, E. (2015). Bullet physics simulation. In ACM timization. arXiv:1703.03161.
SIGGRAPH Courses, 1. Association for Computing Van Nguyen, T.T., Phung, M.D., and Tran, Q.V. (2017b).
Machinery. Behavior-based navigation of mobile robot in unknown
De Silva, L. and Ekanayake, H. (2008). Behavior-based environments using fuzzy logic and multi-objective op-
robotics and the reactive paradigm a survey. In 11th timization. International Journal of Control and Au-
International Conference on Computer and Informa- tomation, 10(2), 349–364.
tion Technology, 36–43. doi:10.1109/ICCITECHN.2008. Ye, C. and Webb, P. (2009). A sub goal seeking approach
4803107. for reactive navigation in complex unknown environ-
Dhaouadi, R. and Hatab, A.A. (2013). Dynamic mod- ments. Robotics and Autonomous Systems, 57(9), 877–
elling of differential-drive mobile robots using lagrange 888.
and newton-euler methodologies: A unified framework.
Advances in Robotics & Automation, 2(2), 1–7.
Horiuchi, Y. and Noborio, H. (2001). Evaluation of
path length made in sensor-based path-planning with