An automated guided logistics robot for

pallet transportation
Rui Lin, Haibo Huang and Maohai Li
School of Mechanical and Electrical Engineering, Soochow University, Suzhou, China

Abstract
Purpose – This study aims to present an automated guided logistics robot mainly designed for pallet transportation. Logistics robot is compactly
designed. It could pick up the pallet precisely and transport the pallet up to 1,000 kg automatically in the warehouse. It could move freely in all
directions without turning the chassis. It could work without any additional infrastructure based on laser navigation system proposed in this work.
Design/methodology/approach – Logistics robot should be able to move underneath and lift up the pallet accurately. Logistics robot mainly
consists of two sub-robots, like two forks of the forklift. Each sub-robot has front and rear driving units. A new compact driving unit is compactly
designed as a key component to ensure access to the narrow free entry of the pallet. Besides synchronous motions in all directions, the two sub-
robots should also perform synchronous lifting up and laying down the pallet. Logistics robot uses a front laser to detect obstacles and locate itself
using on-board navigation system. A rear laser is used to recognize and guide the sub-robots to pick up the pallet precisely within 6 5mm/1o in x-/
yaw direction. Path planning algorithm under different constraints is proposed for logistics robot to obey the traffic rules of pallet logistics.
Findings – Compared with the traditional forklift vehicles, logistics robot has the advantages of more compact structure and higher expandability. It can
realize the omnidirectional movement flexibly without turning the chassis and take zero-radius turn by controlling compact driving units synchronously.
Logistics robot can move collision-free into any pallet that has not been precisely placed. It can plan the paths for returning to charge station and charge
automatically. So it can work uninterruptedly for 7
24 h. Path planning algorithm proposed can avoid traffic congestion and improve the passability of
the narrow roads to improve logistics efficiencies. Logistics robot is quite suitable for the standardized logistics factory with small working space.
Originality/value – This is a new innovation for pallet transportation vehicle to improve logistics automation.
Keywords Automated guided, Logistics robot, Pallet transportation, Driving unit, Path planning, Global localization
Paper type Research paper

1. Introduction and goods around a smart factory or warehouse. As forklift AGVs

use a traditional forklift structure, the overall size and its self-weight
Logistics automation is becoming more important for the smart are both large. A single steering wheel or a differential fixed double
factory (Jeremie, 2017). Logistics inside manufacturing steering wheel with a front integrated steering motor are generally
warehouse mainly includes the transportations of raw materials, adopted as the driving wheels and the two-forked casters with two
semi-finished products and finished goods. Pallet is a basic unit of auxiliary wheels. The arrangement of driving wheels and attached
factory logistics for the products during transportation and storage wheels determines that this type of AGV cannot achieve zero-radius
solutions. Pallet logistics plays a critical role in the improvement of rotation. It also requires a rather large working space.
logistics efficiency and reduction of logistics cost. Traditional The forklift AGVs with special mechanical structure have been
manual hydraulic forklifts were the main choice for pallet delivery developed to transport pallets, such as Agumos G130 (Melkus,
and required a lot of labors. In recent years, more and more 2016) of Melkus Mechatronic GmbH and Nipper (Nipper, 2020)
automation solutions are provided for pallet transportation to of Dutch F3 Design. Agilox developed by INTREST Services
improve logistics efficiency and lower labor costs. GmbH of Germany can run in all directions and perform
The researchers and companies have improved the traditional palletizing automatically (Agilox, 2020). However, it has an over
forklifts based on autonomous navigation technology, called forklift 350 kg self-weight. Picking-up of the pallet is implemented by
automated guided vehicle (AGV) mainly used for pallet logistics (Li double scissors mechanism. These forklift AGVs adopt similar
et al., 2015). Forklift AGV can plan its collision-free path in the compact driving design. To meet the requirement of logistics
factory (Vivaldini et al., 2009). It can transport the pallets automation, logistics robot should have an ability of autonomous
autonomously and intelligently to meet the unmanned logistics navigation. It can move freely between different production lines
requirements. The forklift companies, such as Dematic (Dematic, avoiding obstacles and estimate its accurate pose synchronously.
2020) and Jungheinrich (Jungheinrich, 2019), have developed These logistics robots are developed based on different navigation
forklift AGVs which have been widely used to transport materials technologies in the smart factory.

The current issue and full text archive of this journal is available on Emerald The financial supports by National Key Research and Development Program
Insight at: https://www.emerald.com/insight/0144-5154.htm of China under Grant No. 2019YFB1310004 and National Natural Science
Foundation of China under Grant No. 61973336 are acknowledged.

Received 1 April 2020

Assembly Automation Revised 29 July 2020
© Emerald Publishing Limited [ISSN 0144-5154] 8 September 2020
[DOI 10.1108/AA-04-2020-0052] Accepted 19 October 2020
 Automated guided logistics robot Assembly Automation
Rui Lin, Haibo Huang and Maohai Li

Now the most popular logistics robots used in the actual This paper introduces an automated guided logistics robot
logistics are as follows: mainly used for pallet transportation. A new compact driving unit
 AGV based on magnetic strip navigation (AGV, 2020); is designed especially for logistics robot. It has an advantage of
 forklift AGV based on reflector navigation (Michelson and small size, light weight and omnidirectional locomotion. Logistics
Jull, 1995); robot mainly consists of a front head and two sub-robots. Each
 warehouse robot based on QR code visual navigation sub-robot is similar to a fork of the forklift and has front and rear
(Kiva, 2020 systems); driving units. The sub-robots can move into the narrow free entry
 mobile robot based on natural landmark navigation of the pallet smoothly and accurately. They can synchronously
(Adept, 2020; Chen et al., 2019); and achieve lifting-up and laying-down the pallets by rotating around
 mobile robot based on onboard stereo camera (Sun et al., their vertical center axis. They could transfer pallets autonomously
2019). and intelligently without any human intervention. The front head
of logistics robot mainly has front laser, lithium battery, charging
Global path planning is a key technology for mobile robot system
port, control panel, on board controller and so on. The rear laser
in the logistics application. Each mobile robot can autonomously
mounted backwards is used to detect and recognize the entry of
navigate and travel along with the pre-allocated paths, passing
the pallet, then attain the relative position of the pallet to the sub-
through pre-defined via-points. It can reduce congestion and
robots of logistics robot. The front laser is mainly used for
keep road passability to improve logistics efficiencies and lower
autonomous navigation. For path planning, a 2D grid map is used,
costs. Compared with other methods using magnetic strips and
in which each map cell contains multidimensional criteria.
QR code, mobile robot can navigate freely based on 2D laser
Logistics robot integrates internal odometry and 2D laser range
sensor using natural landmark. In the coexistence (hazardous and
date to get optimal pose estimates by Monte Carlo Localization
populated) environment of workshops and factories, the space
based on Kullback–Leibler divergence.
requirements are smaller and the movements are more flexible.
This paper is organized as follows. In Section 2, a compact
The autonomous navigation technology based on natural
driving unit is introduced. Section 3 discusses logistics robot,
landmarks using laser scanners has developed greatly and been
including its mechanical structure, odometry model and
used widely in recent year. A lot of research work has been carried
motion control. Section 4 presents autonomous navigation
out to solve autonomous navigation based on laser sensors for
system, including path planning algorithm under different
mobile robots, including localization (Boizot et al., 2010), path
constraints and global localization. Section 5 performs several
planning (Connell and La, 2017) and map construction (Thrun,
experiments and analysis to evaluate performance of logistics
2004) algorithms. Robot Operating System (ROS) is the most
robot based on the actual pallet transportation. Finally, Section
famous open source software framework used for navigation
6 deals with the conclusion about our work.
application (ROS, 2019). ROS can achieve map building,
localization and path planning for mobile robot by using only
laser scanner and odometry system. First, ROS can build the 2. Compact driving unit
environment map using the gmapping package. Then mobile Logistics robot should recognize and locate the pallet precisely,
robot can localize itself and receive motion commands, like then achieve lifting-up the pallets and transferring the pallets,
translational velocity and rotational velocity, to run from current finally laying-down the pallets. It should move underneath the
location to a goal location within the global grid map. Mobile pallets freely and smoothly. The driving structure and motion
robot can synchronously avoid moving obstacles during its model should be considered and analyzed carefully because of
navigation and arrive at the destination precisely. small size of the modular unit. It also has ability to lift 1,000 kg
Logistics robot should also have ability to locate and pick-up pallet up and run in all directions.
pallets precisely. Pallet pose estimation with LIDAR sensor and Each logistics robot have two sub-robots and each robot has
Vision is proposed to solve the problem of identifying a pallet, front and rear two driving units. Each driving unit has two DC
which pose is known with large uncertainty (Bellomo et al., brushless motors. To calculate motor performance, we should
2009). Two approaches based on laser scanners are presented consider two different motions: transferring the pallet and
to locate and pick-up pallets with the aim of automating forklift lifting-up the pallet.
trucks (Lecking et al., 2006). The first approach uses reflectors Logistics robot should overcome some resistance during
to calculate the position and orientation of a pallet whereas the transferringXthe pallet. Regardless of air resistance, the
second approach uses only geometrical characteristics. Pallet resistance F is expressed by:
pose estimation of body and fork pose estimation in the pallet X
frame is proposed based on reliable visual feedback for a F ¼ Ff 1 Fi 1 Fj (1)
manipulator with limited degrees of freedom and differential
constraint mobility together with large dimensions of the where Ff is tolling resistance; Fi is climbing resistance; and Fj is
machine (Mohammad et al., 2013). The method introduced to acceleration resistance. Tolling resistance for a sub robot is
integrate the vision output with odometry and realizes smooth expressed by Ff = m mg. m is coefficient of tolling resistance. m
and nonstop transition from global navigation to visual represents its own weight and maximum payload of each sub
servoing. Mohammad and Reza et al. (Mohammad et al., 2016) robot. Because of surface roughness, climbing resistance of the
propose a robust, accurate solution based on fiducial markers sub robot is expressed by Fi = mg sin a. The resistance during
for object manipulation in unstructured outdoor environments acceleration of the sub robot is expressed by Fj = ma.
and unfavorable weather conditions, which relies solely on a As each unit has two driving motors, the driving moment of
monocular camera for pallet detection. every motor is as follows:
 Automated guided logistics robot Assembly Automation
Rui Lin, Haibo Huang and Maohai Li
X
g F  Dw Slip ring is assembled to simplify wiring while two wheels
Md ¼ (2) of the compact driving unit rotate around their vertical
4
rotational axis. Slip ring consists of a stator and a rotor.
where g is factor of safety; and Dw is the radius of driving wheel. The stator is fixed with the shell cover and it keeps
In another case, the maximum power requirement for motionless relative to trapezoidal screw. The rotor rotates
logistics robot is to lift the pallet and maximum payload up. along with the nut by mechanical constraints of two
While two driving wheels integrated in one driving unit rotate guiding axles while two driving wheels rotate. A multi-loop
clockwise around their vertical rotational axis, nut will be absolute encoder as an integrated part of slip ring offers an
driven to rotate synchronously. At the same time, trapezoidal accurate deviation angle of the driving wheels relative to
screw is driven to move upwards to lift the pallet up. the sub robot.
The driving torque required for nut is T = mgP/2p h 0 . P is Compared with hydraulic lifting modules, the driving unit
lead of trapezoidal screw; h 0 is efficiency of trapezoidal lead has the advantages of more compact structure, higher
screw. Assuming the driving force of each wheel is expressed by expandability and safety. The payload weight ratio is quite
Ff = T/2LW and LW is the half distance between two driving high and each driving unit only weighs 3.8 kg; the driving
wheels. Then the motor moment is calculated by the following: unit has a small size and the forklift can move underneath
the pallet; it can hold up payload without circuit breaker and
g 0  Ff  Dw =2
Ml ¼ (3) be safe without power supply; and it can move freely in all
i1  i2  h 1  h 2 directions without turning the chassis and take zero-radius
turn.
where g 0 is factor of safety; i1 and i2 are reduction ratio of
planetary reducer and timing belt; h 1 and h 2 are their
3. Logistics robot
respective efficiency.
Finally the minimal driving torque required for each DC 3.1 Mechanical structure
brushless motor is as follows: Logistics robot consists of two sub robots which are
connected by the head of the robot. The head of the robot
Mm ¼ maxðMl ; Md Þ (4) consists of front laser, lithium battery, charging port, control
panel and on board controller. The front laser is mainly used
Each sub robot integrates the front and rear driving units. A for autonomous navigation. The rear laser mounted
driving unit consists of two DC brushless motors, planetary backwards is used to detect and recognize the entry of the
reducers, rotary encoders, polyurethane wheels, slip ring, pallet, then attain the relative position of the pallet to the
trapezoidal screw, nut, support base and stator. The electrical sub-robots of logistics robot. Mechanical diagram of
interface is also very simple. The compact driving unit is shown logistics robot is shown in Figure 2.
in Figure 1.
The centerline of trapezoidal screw is also the vertical 3.2 Analysis of motion models
rotational axis of two driving wheels. The compact driving unit Motion models of logistics robot are mainly based on the
meets design strength requirements by transient dynamics driving units’ motion. Each driving unit should dynamically
analysis based on ANSYS software. The thread angle of single adjust the relative speed and position. The four driving units
thread nut is as follows: implement synchronous actual motions, such as linear
P 0 f motion, oblique linear motion, zero-radius turn and lifting-
l ¼ arctan < r ¼ arctan (5) up and laying-down. The motion model is shown in
p d2 cosða=2Þ
Figure 3.
where P is the lead of single thread nut; a is lead angle; d2 is
pitch diameter; f is a steel and bronze coefficient of friction; and 3.3 Odometry model
r 0 is equivalent friction angle. As l < r 0 , trapezoid screw can Logistics robot consists of four independent driving units and
hold up load without circuit breaker when nut stops rotating. each driving unit is driven by two DC brushless motors. The
mobile robot can move freely in all directions and take zero-
radius turn. An odometry mainly provides a real-time rough
Figure 1 Three dimensional sketches of the compact driving unit
pose estimate by counting the rotation number of the
stator incremental encoder of each wheel. The accurate deviation
slip ring angle d mi of each driving wheels relative to mobile robot can be
support base trapezoidal given by multi-loop absolute encoder installed inside slip ring of
screw
timing belt 2
screw nut
each driving unit. Figure 4 shows motion model of logistics
planetary guiding axle robot.
reducer 2
wheel 2 motor 2 If d m1,k–1  d m2,k–1  d m3,k–1  d m4,k–1, the odometry [pk qk
driver 2 u k]T at current time k is estimated by:
1 X4
1 2p DW X 8
d k1 ¼  d mi;k1 ; Dsk1 ¼   Dpi;k1
4 i¼1 8 Pe i¼1
(6)
 Automated guided logistics robot Assembly Automation
Rui Lin, Haibo Huang and Maohai Li

Figure 2 Mechanical diagram of logistics robot

Figure 3 Analysis of motion model Figure 4 Motion model of logistics robot

If d m1,k–1 1 d m2,k–1  0, d m3,k–1 1 d m4,k–1  0 and LR · abs(ctan

d m1,k–1 – ctan d m3,k–1)  WR, the odometry [pk qk u k]T at
current time k is estimated by:
1
d 1;k1 ¼  ðd m1;k1  d m2;k1 Þ (8)
2

1
d 2;k1 ¼  ðd m3;k1  d m4;k1 Þ (9)
2

2p DW X 4
Ds1;k1 ¼  Dpi;k1 (10)
8 4  Pe i¼1
>
> pk ¼ pk1 1 Dsk1  cosðu k1 1 d k1 Þ
>
<
qk ¼ qk1 1 Dsk1  sinðu k1 1 d k1 Þ (7) 2p DW X 8
>
> Ds2;k1 ¼  Dpi;k1
>
: 4  Pe i¼5
(11)
u k ¼ u k1

where Dpi,k–1 is incremental changes in position measured by

Ds1;k1  sinDs1;k1 1 Ds2;k1  sinDs2;k1
incremental encoder i of each motor; Pe is the total number of Du k1 ¼ (12)
2  LR
incremental encoder of each motor.
 Automated guided logistics robot Assembly Automation
Rui Lin, Haibo Huang and Maohai Li

Ds1;k1  cosDs1;k1 1 Ds2;k1  cosDs2;k1 where I is a 8
8 identity matrix; and WR is the distance
Drk1 ¼ (13)
2  Du k1 between the rotating axis of driving unit and the center of
parallel mobile robot on X-axis. The following constraint
8   conditions should satisfy:
>
< pk ¼ pk1 1 Drk1   sinðu k1 1 d k1 Þ  sinðu k1 Þ 
> If v r,k = 0, then:
qk ¼ qk1  Drk1  cosðu k1 1 d k1 Þ  cosðu k1 Þ 8
>
>
: u ¼u 1d >
> vr;k þ WR  wr;k
k k1 k1 > d m1;k ¼ d m2;k ¼ arc ctan
< Lw  wr;k
(14) (17)
>
> vr;k  WR  wr;k
>
: d m3;k ¼ d m4;k ¼ arc ctan
where Drk–1 is radius of circular arcs of motion; and LR is the Lw  wr;k
distance between the rotating axis of driving unit and the center
of logistics robot on Y-axis. If v r,k = 0, then:
The internal odometer is widely used to provide easy
accessible real-time rough pose estimation for logistics robot.
d m1;k ¼ d m2;k ¼ d m3;k ¼ d m4;k (18)
The pose errors will accumulate constantly and orientation
errors dominate because they can grow without bound into
translational position errors. Usually the gyroscope is used to 4. Autonomous navigation
correct the orientation error. The position transformation Logistics robot should move to the goal location accurately and
between the consecutive time is applied and as an input then load or unload the pallet autonomously. So a key
parameter in global localization procedure as shown in requirement is the high localization and positioning accuracy of
subsection 4.2. logistics robot especially in industrial applications. While
transferring the pallets in the smart factory, logistics robot
3.4 Motion control should plan a collision-free optimal path to the goal location
Dynamic window approach (DWA) (Foxt et al., 2002; Marija and localize itself in the global grid map. We developed the
and Ivan, 2007) is performed to search optimal translational and whole control software system for autonomous navigation of
rotational velocities [ r,k v r,k]T under mobile robot’s kinematic logistics robot. A is used for global path planning (Nannicini
and dynamic constraints. DWA considers periodically only a et al., 2008), and DWA is performed to search optimal
short time interval when computing the next control command to
translational and rotational velocities under robot’s kinematic
reduce the enormous complexity of the general motion planning
and dynamic constraints. Mobile robot often moves along
problem. Hence, by assuming piecewise constant velocities, we
paths to its goals that it preplans in advance. The optimal poses
can approximate the trajectory of logistics robot by a sequence of
are estimated by Monte Carlo localization (Dieter et al., 1999)
circular and straight line arcs.
based on Kullback–Leibler divergence (Dieter, 2002) in the
While v r,k = 0, mobile robot will move in a straight or
pre-built global occupancy grid map.
oblique linear direction during the next time interval. While
v r,k = 0, mobile robot will move in a circle with center at the
4.1 Path planning under different constraints
point O. The point O that mobile robot rotates about is
For path planning, we use a 2D grid map, in which each map
known as the instantaneous center of curvature (ICC). The
velocity vector  m,k of each driving motor can be calculated cell contains multidimensional criteria relevant for motion,
by: such as the presence of an obstacle, or membership in a
recognized path. These criteria mare merged into a scalar cost
1 ^ m;k
vm;k ¼ IV (15) map, from which an optimal global plan is computed at a 10 Hz
DW rate using a gradient method. Another unique aspect of the
2 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 3 planning method is its use of intermediate line goals, in which
6 ð r;k 1 WR  v r;k Þ 1 LR  v r;k 1 LW  v r;k 7
2 2 2 all points on a line goal are considered equally good. Line goals
6 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7 help to keep the planner from always pushing directly toward
6 7
6 ð r;k 1 WR  v r;k Þ2 1 L2R  v 2r;k  LW  v r;k 7 the goal, when a promising path is found that leads in the
6 7
6 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7 general direction of the goal.
6 7
6 ð r;k 1 WR  v r;k Þ2 1 L2R  v 2r;k 1 LW  v r;k 7 A search algorithm is widely used in path planning and
6 7
6 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7 graph traversal, which is the process of finding a path between
6 7
6 ð r;k 1 WR  v r;k Þ2 1 L2R  v 2r;k  LW  v r;k 7 multiple points, called “nodes.” Specifically, A selects the path
^ m;k 6 7
V ¼ 6 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7
6 7 that minimizes f(n) = g(n) 1 h(n), where n is the next node on
6 ð r;k  WR  v r;k Þ2 1 L2R  v 2r;k 1 LW  v r;k 7
6 7 the path, g(n) is the cost of the path from the start node to n,
6 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7
6 7 and h(n) is a heuristic function that estimates the cost of the
6 ð r;k  WR  v r;k Þ 1 L2R  v 2r;k  LW  v r;k 7
2
6 7 cheapest path from n to the goal. A terminates when the path it
6 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7
6  v r;k 7
6 ð r;k  WR  v r;k Þ 1 L2R  v 2r;k 1 LW
2 chooses to extend is a path from start to goal or if there are no
7
6 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7 paths eligible to be extended. Each iteration of the A algorithm
4 5
ð r;k  WR  v r;k Þ2 1 L2R  v 2r;k  LW  v r;k will be extended from the node with the minimal cost function
(for 4- and 8-connectivity) to the goal. The node whose value is
(16)
253 or 254 will not be extended.
 Automated guided logistics robot Assembly Automation
Rui Lin, Haibo Huang and Maohai Li

While logistics robot works in the actual factory, it should landmarks shown the pre-built global occupancy grid map. Blue
obey some traffic rules just like self-driving care running on a discrete points are the laser points detected by the front lase. The
road of a city. We build a set of special rules to limit the server two orthogonal dotted lines denote the ruler. The purple line
to choose the constraint route nodes for logistics robot. These segments indicate the path planned that will drive the robot from
special rules are configured as Forbidden Zone, Forbidden Line, the start to the goal. Figure 5(a) and (b) shows the results of path
Resistance Zone, Keep Right Zone, Unidirectional Zone and so on. planning while no special zone is set. The path may be the shortest
The first three zones can be achieved by changing the value for one to the goal. We set the value of grid cell to 253 for Forbidden
every grid cell inside the zones. The last two zones can be Zone and 254 for Forbidden Line. The path planned cannot travel
realized by reducing expansion of grids. through Forbidden Zone and Forbidden Line that are also NO GO
Figure 5 shows the results of path planning under different zones. Forbidden Zone is expressed by the pink rectangle as shown
constraints for logistics robot from the current position to the goal in Figure 5(c) and (d). Forbidden Line is expressed by the red line
under the different rules. Logistics robot is marked by a cadetbule as shown in Figure 5(e) and (f). The value of grid cell is set from 1
ellipse. Logistics robot moves between point A and point B to 252. Figure 5(g) and (h) shows the results of path planning
marked by blue rectangles. Black discrete points are the under different resistance values marked by the brown rectangle.
The path planning algorithm will select a path with the minimal
Figure 5 Results of global path planning under different constraints cost f(n). Keep Right Zone is configured by the blue rectangle and
the results are shown in Figure 5(i) and (j). Keep Right Zone is
suitable for configuration of crossroads and busy roads in the
multiply robot system. It assumes that mobile robot can see the
centerline of Keep Right Zone and consider it as a virtual obstacle.
The grid cells are set to bigger values by expanding the virtual
obstacle to the left. Then a path is planned along the right side of
the virtual obstacle. Unidirectional Zone is expressed by dark brown
rectangle and logistics robot is limited to move in only one preset
direction in Unidirectional Zone. The path results are shown in
Figure 5(k) and (l).

4.2 Global localization using the Kullback-Leibler

Divergence-Monte Carlo Localization (KLD-MCL)
algorithm
The global localization method is achieved based on the pre-built
global occupancy grid map. Logistics robot integrates internal
odometry and 2D laser range date to get optimal pose estimates
by Monte Carlo localization based on Kullback–Leibler
divergence. KLD-MCL method is used to obtain an accurate
pose estimate based on the pre-built global occupancy grid map.
It can eliminate the accumulated errors of pose estimate offered
only by internal odometry. We can obtain the optimal pose of
logistics robot by KLD-sampling with adaptive sample set size.
Finally the KD tree is clustered by K-means algorithm and used
to select a set of particles with maximum average weight and
calculate their weighted average as a final optimal pose estimate.
From the pseudo-code of KLD-sampling MCL with
adaptive sample set size as listed in Table 1, we can obtain the
robot pose and its covariance, which needs to be propagated to
the obstacle 2 D position uncertainty. The global position
estimation based on global occupancy grid map is [p q u ]T, and
2 3
s 2p s 2pq s 2pu
6 7
its covariance matrix is QR ¼ 4 s 2pq s 2q s 2qu 5. The
2 2 2
s pu s qu s u
function measurement_score is used to estimate the positioning
score at the current time. The 2D front laser of logistics robot
can get the 541 laser readings per scan. The LIDAR sensor
makes observations by means of 541 laser beams corresponding
to the 541 angles. The laser readings matched with grid in the
pre-built map m can be counted. As we know, if all the 541 laser
readings are matched, the final position calculated can be taken
as the best estimate. Matrix QR is a 3
3 symmetric semi-
definite matrix and represented as an ellipsoid. The projection
 Automated guided logistics robot Assembly Automation
Rui Lin, Haibo Huang and Maohai Li

Table 1 Global localization using the KLD-MCL algorithm

1: MCL_GLOBALLOCALIZATION( x t–1, m t,nzt, m, « , d ) o
½i ½i
2: Input: the previous sample set x t1 ¼ hxt1 ; v t1 iji ¼ 1; . . . ; n , the global map m, the most recent control m t, the measurement zt, the
statistical error bounds « and d and a user-defined minimum of particles M x min
3: x t–1 = Ø
4: M = 0, M x = 0, k = 0, a = 0
5: for all b in H do
6: b = empty
7: endfor
8: do
½i
9: draw i with probability 1 v t1
½M ½i
10: xt =sample_motion_model ( m t, v t1 )
½M ½M
11: v t =measurement_model (zt, xt , m)
½ 
12: a ¼ a 1 v tM
½  ½ 
13: x t ¼ x t1 1 hxtM ; v tM i
½M
14: if xt falls into empty bin b then
15: k=k 1 1
16: b = non – empty
17: if k > 1 then 8 93
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
< =
k1 2 2
18: Mx :¼ 1 1 z1d
2« : 9ð k  1Þ 9ð k  1Þ ;
19: endif
20: endif
21: M=M 1 1
22: while M < M x or M < M x min
23: for i = 1,. . .., n do
½ ½
24: weight normalization:v ti ¼ v ti =a
25: update: x t
26: insert x t into KD tree
27: endfor
28: Cluster KD tree: kdtree_cluster ()
29: Initialize the set of particles to zero
30: Cluster the KD tree by K-means algorithm
31: Select a set of particles with maximum average weight and calculate their weighted average as final pose estimate [p q u ]T with the pose
uncertainty QR
32: c score =measurement_score([p q u ]T, zt, m)

on the global coordinate system axis is an ellipse. Error ellipses lost mobile robot needs human assisted localization to re-
covering one standard deviation in either sides of X and Y localize itself to assure its matching score is greater than a preset
directions are shown. The ellipses cover two standard value.
deviations in either sides of X and Y directions. The projection
of its top view on the global coordinate system axis and axis is 5. Experiment and analysis
an ellipse. The error of mobile robot may be propagated to the
obstacle 2D position uncertainty. Incorporating the following Extensive experiments were conducted to evaluate
uncertainty analysis, the uncertainties for obstacles closer to the performance of logistics robot. The position repeat accuracy
robot tend to be lower, as expected for obstacles with larger and the actual transportation of the pallet were tested for
disparities. logistics robot respectively. An accompanying video shows a
While mobile robot is kidnapped to an arbitrary position, its run of the experiments.
accurate pose may be not recovered by above localization To test position repeat accuracy of logistics robot’s global
method. In other situation, mobile robot is almost surrounded localization, a QR-code was pasted on the rotation center of
by dynamic obstacles, mobile robot may be lost because of the logistics robot and a downward QR code reader was applied to
failure based on above localization method. We can calculate track its real position and direction at a preset point. Logistics
matching score by matching the current laser range date to the robot moved from Point A to Point B repeatedly controlled by
landmarks in the global grid map based on the accurate pose the navigation system. Point B was located under the
estimate. The matching score is taken as confidence level to downward QR code reader. Point A was a random point about
judge the accuracy of pose estimate at the present time. The 10 meters away from Point B. The experiment was repeated
matching score can be used to detect the occurrence of 200 times independently. The 200 positions [p q u ]T relative to
the kidnapped or lost robot problem. While matching score is the center of QR code reader were obtained. The measure data
smaller than a threshold, mobile robot is judged to get lost. The is shown in Figure 6. The average positioning error is 11 mm,
 Automated guided logistics robot Assembly Automation
Rui Lin, Haibo Huang and Maohai Li

Figure 6 Position repeat accuracy tests Figure 7(a) shows logistics robot received a task of pallet
transportation. Then it planned a collision-free optimal path to
the goal location; Figure 7(b) shows logistics robot moved
accurately to the goal location in the global map. The goal
location was a point located directly in front of the pallet;
Figure 7(c) shows logistics robot rotated 180 degree and kept
back to the pallet. Logistics robot detected and obtained the
relative pose of pallet by the rear 2D laser scanner and then
adjusted its pose to assure that each sub robot faced to the free
entry of the pallet directly and precisely; Figure 7(d) shows
logistics robot was guided to move back toward the pallet
guided by rear laser; Figure 7(e) shows logistics robot moved
smoothly through the pallet openings; Figure 7(f) shows the
four driving units were controlled consistently and
simultaneously to lift the pallet up; Figure 7(g) shows logistics
robot transferred the pallet to the final location and laid the
pallet down; Figure 7(h) shows logistics robot left the pallet and
robot arrived at the standby point front of the pallet.
Figure 8 shows a grid map of the working scene and real
trajectory of logistics robot. The global grid map consisted of
grid nodes and pre-set traffic constraints were expressed by
lines and rectangles with different colors. The map was stored
in the on-board industrial computer. The motion paths of

Figure 7 Entire workflow of the pallet transportation

and the maximum error is 26.18 mm. 42.57% is smaller than

10 mm, and 94.55% is smaller than 20 mm.
Logistics robot designed in our work consisted of two forks
and a main header. It integrated four compact driving units. It
used a front 2D scanning laser to navigate freely in the working
scenes. It could get its optimal pose estimate and avoid moving
and temporary obstacles during pallet delivery. It used rear
laser to detect and attain the relative position of the pallet which
may be not precisely placed. We set up this experiment to
design and tested a pallet transportation task in which mobile
robot had to locate, approach, obtain the pose and finally pick
up the pallet.
The entire workflow of pallet transportation composed of a
sequence of 10 pictures is shown in Figure 7. Logistics robot
transported the pallet from Pallet Point to Park Point and then
returned back to Pallet Point. Logistics robot could arrive at the
front of the pallet and then pick-up the pallet precisely and
safely. The pallet could also be transported to a specified
location and laid-down accurately and automatically. The
whole process was automatically implemented by simulating an
actual task workflow.
 Automated guided logistics robot Assembly Automation
Rui Lin, Haibo Huang and Maohai Li

Figure 8 Grid map and real trajectory of logistics robot AGV (2020), “Automatic guided vehicles – AGV”, available at:
www.agvsolutions.com/ (accessed 22 March 2020).
Bellomo, N., Marcuzzi, E., Baglivo, L., Pertile, M., Bertolazzi,
E. and De Cecco, M. (2009), “Pallet pose estimation with
LIDAR and vision for autonomous forklifts”, Proceedings of
the 13th IFAC Symposium on Information Control Problems in
Manufacturing, June 3-5, Moscow, pp. 612-617.
Boizot, N., Busvelle, E. and Gauthier, J.P. (2010), “An
adaptive high-gain observer for nonlinear systems”,
Automatica, Vol. 46 No. 9, pp. 1483-1488.
Chen, H., Huang, H.L., Qin, Y. and Liu, Y.H. (2019), “Vision
and laser fused SLAM in indoor environments with Multi-
Robot system”, Assembly Automation, Vol. 39 No. 2,
pp. 297-307.
Connell, D. and La, H.M. (2017), “Dynamic path planning
logistics robot, which consisted of discrete position points, were and replanning for mobile robots using RRT ”, IEEE
stored in a LOG file in real time. The position points were International Conference on Systems, Man and Cybernetics
sampled from the LOG file every 10 points to show more (SMC), 1-4 October, Banff, pp. 1429-1434.
clearly and simply. Each position point was a vector with x and Dematic (2020), “KION GROUP AG j homepage”, available
y coordinates and direction in the global map. The purple line at: www.kiongroup.com/en/ (accessed 17 March 2020).
segments with an arrowhead denoted a real trajectory during Dieter, F. (2002), “KLD-Sampling: adaptive particle filters”,
the transportation process of logistics robot. The purple line Advances in Neural Information Processing Systems, Vol. 14,
segments only simply showed the trajectory from Pallet Point to pp. 713-720.
Park Point and then back to Pallet Point in this experiment. Dieter, F., Wolfram, B., Dellaert, F. and Sebastian, T. (1999),
Logistics robot also obeyed the pre-set traffic rules, such as “Monte Carlo localization: efficient position estimation for
Forbidden Zone and Unidirectional Zone. mobile robots”, the Sixteenth National Conference on Artificial
Intelligence and Eleventh Conference on Innovative Applications
of Artificial Intelligence, 18-22 July, FL, pp. 343-349.
6. Conclusion
Foxt, D., Burgardt, W. and Thrun, S. (2002), “Controlling
This work presents an automated guided logistics robot mainly synchro-drive robots with the dynamic window approach to
designed for pallet transportation. Logistics robot can pick-up the collision avoidance”, IEEE/RSJ International Conference on
pallet precisely and transport the pallet up automatically in the Intelligent Robots & Systems, 30 September-5 October, Osaka,
warehouse. The robot can move underneath the narrow free pp. 1280-1287.
entry of the pallet easily and accurately. A new compact driving Jeremie, C. (2017), A Look into Logistics Automation, ROBO
unit is compactly designed as a key component to ensure access Global Defining the Universe of Robotics and Automation, ETF
to the narrow free entry of the pallet. It has an advantage of mall Securities, London, pp. 1-18.
size and omnidirectional locomotion. The autonomous Jungheinrich (2019), “Homepage j jungheinrich”, available at:
navigation system is also proposed based on 2D scanning laser www.jungheinrich.co.th/ (accessed 1 March 2020).
using the natural landmarks of working scenes. Logistics robot Kiva (2020), “Kiva systems - Amazon robotics: : We reimagine
uses the front laser sensor to navigate freely and safely based on now”, available at: www.amazonrobotics.com/#/(accessed
the natural landmarks. Logistics robot can move collision-free 22 March 2020).
into any pallet that has not been precisely placed using a rear Lecking, D., Wulf, O. and Wagner, B. (2006), “Variable pallet
laser. It can plan a free path for returning to charge station and Pick-Up for automatic guided vehicles in industrial
charge automatically. Path planning algorithm proposed can environments”, IEEE Conference on Emerging Technologies
and Factory Automation, 20-22 September, Prague,
avoid traffic congestion and improve the passability of the narrow
pp. 1169-1174.
roads to improve logistics efficiencies. So it can work
Li, L.Y., Liu, Y.Y., Fang, M., Zheng, Z.Z. and Tang, H.B.
uninterruptedly for 7
24 h. Compared with the traditional
(2015), “Vision-Based intelligent forklift automatic guided
forklift AGV, logistics robot has the advantages of more compact
vehicle (AGV)”, Proceedings of the 2015 IEEE International
structure and higher expandability. Therefore, the new compact
Conference on Automation Science and Engineering, 24-28
driving unit is quite suitable for logistics robot in the smart factory August, TX, pp. 264-265.
with small working space. Marija, S. and Ivan, P. (2007), “Dynamic window based
approach to mobile robot motion control in the presence of
References moving obstacles”, Proceedings of the 2007 IEEE International
Conference on Robotics and Automation, 10-14 April, Roma,
Adept (2020), “Adept technology inc”, available at: https:// pp. 1986-1991.
robotics.omron.com/home/?region=ch/ (accessed 21 March Melkus, A. (2016), Fahreloses transportfarzeug, European
2020). patent, no. EP3216747A1.
Agilox (2020), “Agilox system”, available at: http://agilox.net/ Michelson, D.G. and Jull, E.V. (1995), “Depolarizing trihedral
(accessed 28 March 2020). corner reflectors for radar navigation and remote sensing”,
 Automated guided logistics robot Assembly Automation
Rui Lin, Haibo Huang and Maohai Li

IEEE Transactions on Antennas & Propagation, Vol. 43, environments with onboard stereo camera for MAVs”, IEEE
pp. 513-518. Transactions on Industrial Informatics, preprinting.
Mohammad, M.A., Reza, G., Antti, K. and Jouni, M. (2016), Thrun, S. (2004), “Simultaneous localization and mapping
“A multistage controller with smooth switching for with sparse extended information filters”, The International
autonomous pallet picking”, 2016 IEEE International Journal of Robotics Research, Vol. 23 Nos 7/8, pp. 693-716.
Conference on Robotics and Automation (ICRA), 16-21 May, Vivaldini, K.T., Caurin, G. and Becker, M. (2009),
Stockholm, pp. 2535-2542. “Automatic routing of forklift robots in warehouse
Mohammad, M.A., Reza, G., Antti, K., Mika, H., Kalevi, H. applications”, 20th International Congress of Mechanical
and Jouni, M. (2013), “Position-Based visual servoing for Engineering, 15-20 November, Gramado, pp. 1-10.
pallet picking by an articulated-frame-steering hydraulic
mobile machine”, 2013 6th IEEE Conference on Robotics,
Automation and Mechatronics (RAM), 12-15 November, Further reading
Manila, pp. 218-224. Young, K.D., Utkin, V.I. and Ozguner, U. (1999), “A
Nannicini, G., Delling, D., Liberti, L. and Schultes, D. (2008), control engineer’s guide to sliding mode control”, IEEE
“Bidirectional a search for Time-Dependent fast paths”, Transactions on Control Systems Technology, Vol. 7 No. 3,
International Workshop on Experimental & Efficient Algorithms, pp. 328-342.
Springer Berlin Heidelberg, Vol. 5038, pp. 334-346. Se, S., Lowe, D. and Little, J. (2001), “Vision-based mobile
Nipper (2020), “F3-Design”, available at: www. robot localization and mapping using scale-invariant
dinostretchhood.com/nipper/ (accessed 25 March 2020). features”, IEEE International Conference on Robotics &
ROS (2019), “ROS.org j powering the world’s robots”, Automation, 21-26 May, Seoul, pp. 2051-2058.
available at: www.ros.org/ (accessed 29 March 2020).
Sun, J.B., Song, J., Chen, H.Y. and Liu, Y.H. (2019), Corresponding author
“Autonomous state estimation and mapping in unknown Haibo Huang can be contacted at: [email protected]

For instructions on how to order reprints of this article, please visit our website:
www.emeraldgrouppublishing.com/licensing/reprints.htm
Or contact us for further details: [email protected]