278 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 7, NO.

3, JUNE 1991

The Vector Field Histogram-Fast Obstacle

Avoidance for Mobile Robots
Johann Borenstein, Member, IEEE, and Yoram Koren, Senior Member, IEEE

Abstract-A new real-time obstacle avoidance method for [5]. While the VFF method provides superior real-time ob-
mobile robots has been developed and implemented. This stacle avoidance for fast mobile robots, some limitations
method, named the vector field histogram (VFH), permits the concerning fast travel among densely cluttered obstacles were
detection of unknown obstacles and avoids collisions while
simultaneously steering the mobile robot toward the target. identified in the course of our experimental work [18]. To
The VFH method uses a two-dimensional Cartesian his- overcome these limitations, we developed a new method,
togram grid as a world model. This world model is updated named vectorfield histogram (VFH), which is introduced in
continuously with range data sampled by on-board range sen- Section IV. The VFH method eliminates the shortcomings of
sors. The VFH method subsequently employs a two-stage data- the VFF method, yet retains all the advantages of its prede-
reduction process in order to compute the desired control com-
mands for the vehicle. In the first stage the histogram grid is cessor (as will be shown in Section IV). A comparison of the
reduced to a one-dimensional polar histogram that is con- VFH method to earlier methods is given in Section V, and
structed around the robot’s momentary location. Each sector in Section VI presents experimental results obtained with our
the polar histogram contains a value representing the polar VFH-controller mobile robot.
obstacle density in that direction. In the second stage, the
algorithm selects the most suitable sector from among all polar 11. SURVEY
OF EARLIER
OBSTACLE-AVOIDANCE
histogram sectors with a low polar obstacle density, and the METHODS
steering of the robot is aligned with that direction.
Experimental results from a mobile robot traversing densely This section summarizes relevant obstacle avoidance meth-
cluttered obstacle courses in smooth and continuous motion and ods, namely edge-detection, certainty grids, and potential
at an average speed of 0.6-0.7 m/s demonstrate the power of field methods.
the VFH method.
A. Edge-Detection Methods
I. INTRODUCTION One popular obstacle avoidance method is based on edge-
detection. In this method, an algorithm tries to determine the
0 BSTACLE avoidance is one of the key issues to suc-
cessful applications of mobile robot systems. All mobile
robots feature some kind of collision avoidance, ranging
position of the vertical edges of the obstacle and then steer
the robot around either one of the “visible” edges. The line
from primitive algorithms that detect an obstacle and stop the connecting two visible edges is considered to represent one of
robot short of it in order to avoid a collision through sophisti- the boundaries of the obstacle. This method was used in our
cated algorithms that enable the robot to detour obstacles. very early research [4], as well as in several other works
The latter algorithms are much more complex, since they [111, [213, [22], [28], all using ultrasonic sensors for obstacle
involve not only the detection of an obstacle, but also some detection. A disadvantage with current implementations of
kind of quantitative measurements concerning the dimensions this method is that the robot stops in front of obstacles to
of the obstacle. Once these have been determined, the obsta- gather sensor information. However, this is not an inherent
cle avoidance algorithm needs to steer the robot around the limitation of edge-detection methods; it may be possible to
obstacle and proceed toward the original target. Usually, this overcome this problem with faster computers in future imple-
procedure requires the robot to stop in from of the obstacle, mentations.
take the measurements, and only then resume motion. Obsta- In another edge-detection approach (using ultrasonic sen-
cle avoidance (also called reflexive obstacle avoidance or sors), the robot remains stationary while taking a panoramic
local path planning) may result in nonoptimal paths [5] since scan of its environment [13], [14]. After the application of
no prior knowledge about the environment is used. certain line-fitting algorithms, an edge-based global path
A brief survey of relevant earlier obstacle avoidance meth- planner is instituted to plan the robot’s subsequent path.
ods is presented in Section II, while Section 111 summarizes A common drawback of both edge-detection approaches is
their sensitivity to sensor accuracy. Ultrasonic sensors pre-
the virtual force field (VFF), an obstacle-avoidance method
developed earlier by our group at the University of Michigan sent many shortcomings in this respect:
Poor directionality limits the accuracy in determining the
Manuscript received July 6, 1989; revised September 1, 1990. This work
spatial position of an edge to 10-50 cm, depending on the
was supported by the U.S. Department of Energy under Grant DE-FGO2- distance to the obstacle and the angle between the obstacle
86NE37969. surface and the acoustic axis.
The authors are with the Advanced Technology Laboratories, The Univer- Frequent misreadings are caused by either ultrasonic
sity of Michigan, Ann Arbor, MI 48109.
IEEE Log Number 9143260. noise from external sources or stray reflections from neigh-

1042-296X/91/06OO-0278$01.OO @ 1991 IEEE

BORENSTEIN AND KOREN: THE VECTOR FIELD HISTOGRAM 279

boring sensors (i.e., crosstalk). Misreadings cannot always

be filtered out, and they cause the algorithm to falsely detect
edges.
Specular reflections occur when the angle between the
wavefront and the normal to a smooth surface is too large. In
this case the surface reflects the incoming ultrasound waves
away from the sensor, and the obstacle is either not detected
or is “seen” as much smaller than it is in reality (since only
part of the surface is detected).
Any one of these errors can cause the algorithm to deter-
mine the existence of an edge at a completely wrong location,
often resulting in highly unlikely paths.
B. The Certainty Grid f o r Obstacle Representation
A method for probabilistic representation of obstacles in a
grid-type world model has been developed at Carnegie-Mel-
Measured
distance d
\ I ,/
lon University (CMU) [13], [23], [24]. This world model,
called a certainty grid, is especially suited to the accommo-
dation of inaccurate sensor data such as range measurements
from ultrasonic sensors.
In the certainty grid, the robot’s work area is represented
by a two-dimensional array of square elements, denoted as
cells. Each cell contains a certainty value (CV) that indi- sensor
cates the measure of confidence that an obstacle exists within
Fig. 1 . Two-dimensional projection of the conical field of view of an
the cell area. With the CMU method, CV’s are updated by a ultrasonic sensor. A range reading d indicates the existence of an object
probability function that takes into account the characteristics somewhere within the shaded region A .
of a given sensor. Ultrasonic sensors, for example, have a
conical field of view. A typical ultrasonic sensor [25] returns Krogh [19] has enhanced this concept further by taking
a radial measure of the distance to the nearest object within into consideration the robot’s velocity in the vicinity of
the cone, yet does not specify the angular location of the obstacles. Thorpe [27] has applied the potential field method
object. (Fig. 1 shows area A in which an object must be to off-line path planning, and Krogh and Thorpe [20] suggest
located in order to result in a distance measurement d ) . If an a combined method for global and local path planning, which
object is detected by an ultrasonic sensor, it is more likely uses a “generalized potential field” approach. Newman and
that this object is closer to the acoustic axis of the sensor Hogan [15] introduce the construction of potential functions
than to the periphery of the conical field of view [4]. For this through combining individual obstacle functions with logical
reason, CMU’s probabilistic function C , increases the CV’s operations.
in cells close to the acoustic axis more than the CV’s in cells Common to these methods is the assumption of a known
at the periphery. and prescribed world model, in which simple predefined
In CMU’s applications of this method [23], [24], the geometric shapes represent obstacles and the robot’s path is
mobile robot remains stationary while it takes a panoramic generated off-line.
scan with its 24 ultrasonic sensors. Next, the probabilistic While each of the above methods features valuable re-
function C , is applied to each of the 24 range readings, finements, none have been implemented on a mobile robot
updating the certainty grid. Finally, the robot moves to a new with real sensory data. By contrast, Brooks [8], [9] and
location, stops, and repeats the procedure. After the robot Arkin [ 11 use a potential field method on experimental mobile
traverses a room in this manner, the resulting certainty grid robots (equipped with a ring of ultrasonic sensors). Brooks’
represents a fairly accurate map of the room. A global implementation treats each ultrasonic range reading as a
path-planning method is then employed for off-line calcula- repulsive force vector. If the magnitude of the sum of the
tions of subsequent robot paths. repulsive forces exceeds a certain threshold, the robot stops,
C . Potential Field Methods turns in the direction of the resultant force vector, and moves
The idea of imaginary forces acting on a robot has been on. In this implementation, however, only one set of range
suggested by Khatib [16]. In this method, obstacles exert readings is considered, while previous readings are lost.
repulsive forces, while the target applies an attractive force to Arkin’s robot employs a similar method; his robot was
the robot. A resultant force vector R , comprising the sum of able to traverse an obstacle course at 0.12 cm/s
a target-directed attractive force and repulsive forces from (0.4 f/s).
obstacles, is calculated for a given robot position. With R as 111. THEVIRTUAL
FORCEFIELDMETHOD
the accelerating force acting on the robot, the robot’s new
position for a given time interval is calculated, and the The virtual force field (VFF) method is our earlier real-
algorithm is repeated. time obstacle avoidance method for fast-running vehicles [ 5 ] .
280 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 7, NO. 3, JUNE 1991

Target
Histogram @
grid

Direction-
-~
of motion

(4 (b)
Fig. 2. (a) Only one cell is incremented for each range reading. with
ultrasonic sensors, this is the cell that lies on the acoustic axis and corre-
sponds to the measured-distance d . (b)A histogramic probability distribution
is obtained by continuous and rapid sampling of the sensors while the vehicle
is moving.
Fig. 3. The virtual force field (VFF) concept: Occupied cells exert repul-
sive forces onto the robot; the magnitude is proportional to the certainty
Unlike the methods reviewed above, the VFF method allows value c i , , of the cell and inversely proportional to d 2 .
for fast, continuous, and smooth motion of the controlled
vehicle among unexpected obstacles and does not require the grid, so the probabilistic sensor information can be used
vehicle to stop in front of obstacles. efficiently to control the vehicle. Fig. 3 shows how this
algorithm works :
A . The VFF Concept As the vehicle moves, a window of w, x w, cells accom-
The individual components of the VFF method are pre- panies it, overlying a square region of C . We call this
sented below. region the “active region” (denoted as C * ) ,and cells that
momentarily belong to the active region are called “active
1) The VFF method uses a two-dimensional Cartesian his- cells” (denoted as c: j ) . In our current implementation,
togram grid C for obstacle representation. As in CMU’s the size of the window is 33 X 33 cells (with a cell size of
certainty grid concept, each cell ( i , j ) in the histogram 10 cm x 10 cm), and the window is always centered about
grid holds a certainty value ci, that represents the confi- the robot’s position. Note that a circular window would be
dence of the algorithm in the existence of an obstacle at geometrically more appropriate, but is computationally
that location. more expensive to handle than a square one.
The histogram grid differs from the certainty grid in the Each active cell exerts a virtual repulsive force Fi,
way it is built and updated. CMU’s method projects a toward the robot. The magnitude of this force is propor-
probability profile onto those cells that are affected by a tional to the certainty value c: and inversely proportional
range reading; this procedure is computationally intensive to d”, where d is the distance between the cell and the
and would impose a heavy time penalty if real-time execu- center of the vehicle, and x is a positive real number (we
tion on an on-board computer was attempted. Our method, assume x = 2 in the following discussion). At each itera-
on the other hand, increments only one cell in the his- tion, all virtual repulsive forces are totaled to yield the
togram grid for each range reading, creating a “probabil- resultant repulsive force Fr. Simultaneously, a virtual at-
ity ’ ’ distribution’ with only small computational over- tractive force Fl of constant magnitude is applied to the
head. For ultrasonic sensors, this cell corresponds to the vehicle, “pulling” it toward the target. The summation of
measured distance d [see Fig. 2(a)3 and lies on the acous- F,. and Ft yields the resulting force vector R . In order to
tic axis of the sensor. While this approach may seem to be compute R , up to 33 x 33 = 1089 individual repulsive
an oversimplification, a probabilistic distribution is actu- force vectors Fi, must be computed and accumulated.
ally obtained by continuously and rapidly sampling each The computational heart of the VFF algorithm is therefore
sensor while the vehicle is moving. Thus, the same cell a specially developed algorithm for the fast computation
and its neighboring cells are repeatedly incremented, as and summation of the repulsive force vectors.
shown in Fig. 2(b). This results in a histogramic proba- 3) Combining concepts 1 and 2 (given above) in real-time
bility distribution in which high certainty values are enables sensor data to influence the steering control imme-
obtained in cells close to the actual location of the obstacle. diately. In practice, each range reading is recorded into
2) Next, we apply the potential field idea to the histogram the histogram grid as soon as it becomes available, and the
subsequent calculation of R takes this data point into
‘We use the term “probability” in the literal sense of “likelihood.” account. This feature gives the vehicle fast response to
BORENSTEIN AND KOREN: THE VECTOR FIELD HISTOGRAM 281

obstacles that appear suddenly, resulting in fast reflexive - 9 e r t a i n t y values

\
behavior imperative at high speeds.

B. Shortcomings of the VFF Method

The VFF method has been implemented and extensively
tested on board a mobile robot equipped with a ring of 24
ultrasonic sensors (see Section VI). Under most conditions,
the VFF-controlled robot performed very well. Typically, it
traversed an obstacle course at an average speed of 0.5 m/s,
provided the obstacles were places at least 1.8 m apart (the \
robot diameter is 0.8 m). With less clearance between two Active cells
~

obstacles (e.g., a doorway), some problems were encoun-

tered. Sometimes, the robot would not pass through a door-
I
way because the repulsive forces from both sides of the
Fig. 4. Mapping of active cells onto the polar histogram.
doorway resulted in a force that pushed the robot away.
Another problem arose out of the discrete nature of the
histogram grid. In order to efficiently calculate repulsive
forces in real time, the robot’s momentary position is mapped vector F,. Hundreds of data points are reduced in one step to
onto the histogram grid. Whenever this position changes only two items: the direction and the magnitude of -Fr.
from one cell to another, drastic changes in the resultant R Consequently, detailed information about the local obsta-
may be encountered. The following numeric example ex- cle distribution is lost.
plains this point. Consider a repulsive force generated by a To remedy this shortcoming, we have developed a new
cell ( m , n) and applied to the robot’s momentary position at method called the vectorfield histogram (VFH). The VFH
+
( m , n 6 ) , which is six cells away (i.e., 0.6 m, with a cell method employs a two-stage data-reduction technique,
size of 10 x 10 cm). The magnitude of this particular force rather than the single-step technique used by the VFF method.
vector is 1 Fm, 1 = k/0.6* = 2.8k. As the robot advances Thus, three levels of data representation exist:
+
by one cell, and its position corresponds to cell ( m , n 5 ) ,
the new force vector is 1 FA, 1 = k/0.5* = 4k. The change 1) The highest level holds the detailed description of the
is 42 % . Changes of this magnitude cause considerable fluc- robot’s environment. In this level, the two-dimensional
tuations in the steering control. The situation is even worse Cartesian histogram grid C is continuously updated in real
when the magnitude of the target-directed constant attractive time with range data sampled by the on-board range sen-
force F, lies between the directions of the two successive sors. This process is identical to the one described in
forces 1 F 1 and 1 F’ I (this condition is, in fact, most likely Section I11 for the VFF method.
because it corresponds to the “steady-state’’ condition when 2) At the intermediate level, a one-dimensional polar his-
the robot travels alongside an obstacle). In this situation, the togram H is constructed around the robot’s momentary
direction of the resultant R may fluctuate by up to 180”. For location. H comprises n angular sectors of width CY. A
this reason it is necessary to smooth the control signal to the transformation (described in Section IV-A below) maps the
steering motor by adding a low-pass filter to the VFF control active region C* onto H , resulting in each sector k
loop [5]. This filter, however, introduces a delay that ad- holding a value h , that represents the polar obstacle
versely affects the robot’s steering response to unexpected density in the direction that corresponds to sector k .
obstacles. 3) The lowest level of data representation is the output of the
Finally, we also identified a problem that occurs when the VFH algorithm: the reference values for the drive and
robot travels in narrow corridors. When traveling along the steer controllers of the vehicle.
centerline between the two corridor walls, the robot’s motion
is stable. If, however, the robot strays slightly to either side The following sections describe the two data-reduction
stages in more detail.
of the centerline, it experiences a strong virtual repulsive
force from the closer wall. This force usually “pushes” the
robot across the centerline, and the process repeats with the A . First Data Reduction and Creation of the Polar
other wall. Under certain conditions, this process results in Histogram
oscillatory and unstable motion [ 171, [ 181.
The first data-reduction stage maps the active region C* of
IV. THEVECTOR FIELDHISTOGRAM METHOD the histogram grid C onto the polar histogram H , as fol-
lows: As with our earlier VFF method, a window moves
Careful analysis of the shortcomings of the VFF method with the vehicle, overlying a square region of ws x w, cells
reveals its inherent problem: excessively drastic data reduc- in the histogram grid (see Fig. 4). The contents of each active
tion occurs when the individual repulsive forces from his- cell in the histogram grid are now treated as an obstacle
togram grid cells are totaled to calculate the resultant force vector, the direction of which is determined by the direction
282 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. I, NO. 3, JUNE 1991

Targe YA Targl
0 n. 0

3'-
...
.
.-
2 2. .I

Partition .
..
..
8 .
..
..
C
..
-. Cutahty V J a r

..
.-
-
dDwn as
-.-
1 1 . CV.1-3
0 cv4-5
is. -
. . .. .... .
SLrt Start cv.1-9

-.
Walls cV=1&12
\Walls cv=l3-15
I A.
1 2 3 M)x 1 2 3 w'w

(a) (b)
Fig. 5. (a) Example of an obstacle course. (b) The corresponding his-
togram grid representation.

0 from the cell to the vehicle centerpoint (VCP)2 360/a is an integer (e.g., a = 5" and n = 72). Each sector
Yj - YO
k corresponds to a discrete angle p quantized to multiples of
p . J. = t a n - 1 - a, such that p = ka, where k = 0, 1,2; -,n - 1. corre-
1,
xi - xo spondence between c: and sector k is established through
and the magnitude is given by
k = INT(P,,,/a). (3)
mi, = (c: j)2( a - bdi,,) (2) For each sector k, the polar obstacle density h is calculated
where by
a, b are positive constants,
is the certainty value of active cell ( i , j ) ,
hk = cm1,J.
1, J
(4)
cr
di, is the histance between active cell ( i , j) and the VCP, E k h active cell is related to a certain sector by (1) and (3).
m i , is the magnitude of the obstacle vector at cell (i, j), In Fig. 4, which shows the mapping from C* into H, all
xo, yo are the present coordinates of the VCP, active cells related to sector k have been highlighted. Note
x i , y j are the coordinates of active cell ( i , j ) , and tliat the sector width in Fig. 4 is a = 10" (not a = 5 " , as in
pi, is the direction from active cell (i, j) to the VCP. the actual algorithm) to clarify the drawing.
Notice that Because of the discrete nature of the histogram grid, the
result of this mapping may appear ragged and cause errors in
1 ) Ct is squared. This expresses our confidence that recur-
the selection of the steering direction (as explained in Section
ring range readings represent actual obstacles, as opposed
IV-B). Therefore, a smoothing function is applied to H,
to single Occurrences of range readings, which may be
which is defined by
caused by noise.
2) mi, is proportional to - d . Therefore, occupied cells
produce large vector magnitudes when they are in the
immediate vicinity of the robot and smaller ones when they 21+ 1
are further away. Specifically, a and b are chosen such (5)
that a - bd,,, = 0 , where d,,, = &(w, - 1 ) / 2 is the
distance between the farthest active cell and the VCP. This where h',k is the smoothed polar obstacle density (POD).
way mi, = 0 for the farthest active cell and increases In our current implementation, I = 5 yields satisfactory
linearly for closer cells. smoothing results.
Fig. 5(a) shows a typical obstacle setup in our lab. Note
H has a arbitrary angular resolution a such that TI = that the gap between obstacles B and C is only 1.2 m and
that A is a thin pole of 3/4-in diameter. The histogram grid
'For symmetrically shaped vehicles, the VCP is easily defined as the obtained after partially traversing this obstacle course is
geometric center of the vehicle. For rectangular vehicles, it is possible to shown in Fig. 5@). The (smoothed) polar histogram corre-
chose two VCP's, e . g . , one each at the center point of the front and rear
axles. sponding to the momentary position of the robot 0 is shown
BORENSTEIN AND KOREN: THE VECTOR FIELD HISTOGRAM 283

Targei

(a) (b)
Fig. 6. (a) Polar obstacle density represented in the smoothed polar
histogram H’(k)-relative to the robot’s position at 0 [in Fig. 5@)]. @)
The same polar histogram as in (a), shown in polar form and overlaying part
of the histogram grid of Fig. 5 @ ) .

further necessary to choose a suitable sector within that

Target valley. The following discussion explains how the algorithm
finds this sector and thus the required steering direction.
I At first, the algorithm measures the size of the selected
tight
sectors =
Candidatedirections
for safe travel
valley (i.e., the number of consecutive sectors with POD’s
POD <threshold below the threshold). Here, two types of valleys are distin-
Unsafe directions,
guished, namely, wide and narrow ones. A valley is consid-
Dark
sector = prohibited for travel ered wide if more than s, consecutive sectors fall below
POD >threshold
thredld the threshold (in our system s, = 18). Wide valleys result
rtition from wide gaps between obstacles or from situations where
C
only one obstacle is near the vehicle. Fig. 8 shows a typical
obstacle configuration that produces a wide valley. The sector
Fig. 7. A threshold on the polar histogram determines the candidate
directions for subsequent travel. that is nearest to kbrg and below the threshold is denoted k ,
and represents the near border of the valley. The fur
border is denoted as k, and is defined as k , = k , s,.+
in Fig. 6(a). The directions (in degrees) in the polar his- The desired steering direction 8 is then defined as 8 = ( k ,
togram correspond to directions measured counterclockwise + k,)/2. Fig. 8 demonstrates why this method results in a
from the positive x axis of the histogram grid. The peaks stable path when traveling alongside an obstacle: If the robot
A , B , and C in the polar histogram result from obstacle approaches the obstacle too closely [Fig. 8(a)], 8 points away
clusters A , B , and C in the histogram grid. Fig. 6(b) shows from the obstacle. If the robot is further away from the
the polar form of the exact same polar histogram as Fig. 6(a), obstacle, 8 allows the robot to approach the obstacle more
overlaying part of the histogram grid of Fig. 5(b). closely [Fig. 8(b)]. Finally, when traveling at the proper
distance d , from the obstacle [Fig. 8(c)], 8 is parallel to the
B. Second Data Reduction and Steering Control obstacle boundary and small disturbances from this parallel
The second data-reduction stage computes the required path are corrected as described above. Note that the distance
steering direction 8. This section explains how 8 is com- d , is mostly determined by smax. The larger s,, the further
puted. the robot will keep from an obstacle, under steady-state
As can be seen in Fig. 6, a smoothed polar histogram conditions.
typically has “peaks,” i.e., sectors with high POD’s, and The second case, a narrow valley, is created when the
“valleys,” i.e., sectors with low POD’s. Any valley com- mobile robot travels between closely spaced obstacles, as
prised of sectors with POD’s below a certain threshold (see shown in Fig. 9. Here the far border k, is less than smaX
discussion in Section IV-C) is called a candidate valley. Fig. sectors apart from k,. However, the direction of travel is
7 visualizes the match between candidate valleys and the again chosen as 8 = ( k , + k f ) / 2 and the robot maintains a
actual environment: Based on the threshold and the polar course centered between obstacles.
histogram of Fig. 6, candidate valleys are shown as lightly Note that the robot’s ability to pass through narrow pas-
shaded sectors in Fig. 7, while unsafe directions (i.e., those sages and doorways results from the ability to identify a
with POD’s above the threshold) are shown in darker shades. narrow valley and to choose a centered path through that
Usually there are two or more candidate valleys and the valley. This feature is made possible through the intermediate
VFH algorithm selects the one that most closely matches the data representation in the polar histogram. Our earlier VFF
direction to the target kbrg (an exception to this rule is method and other potential field methods, by contrast, lack
discussed in Section IV-E). Once a valley is selected, it is this ability [ 181.
284 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 7, NO. 3, JUNE 1991

Fig. 9. Finding the steering reference direction 0 when ,k is obstructed

by an obstacle.

Lowering or raising the threshold even by a factor of 3 or 4

only affects the width of the candidate valleys. This, in turn,
has only a small effect on narrow valleys, since the steering
direction is chosen in the center of the valley. In wide
valleys, only the distance d , from the obstacle is affected.
Severe maladjustments of the threshold have the following
effect on the system performance:
1) If the threshold is much too large [e.g., higher than peak A
in Fig. 6(a)], the robot is not “aware” of that obstacle and
approaches it on a collision course. However, during the
approach, additional sensor readings further increase the
CV’s representing that obstacle, while the distance d to the
affected cells decreases. As is evident from ( 2 ) ,this results
in higher POD’S and consequently in a higher “peak” that
eventually exceeds the threshold. However, in this case the
robot may approach the obstacle too closely (especially
when traveling at high speed) and collide with the objects.
Fig. 8. Obtaining a stable path when traveling alongside an obstacle: (a) 0 2) If, on the other hand, the threshold is much too low, some
points away from the obstacle when the robot is too close. (b) 0 points
toward the obstacle when the robot is further away. (c) Robot runs alongside potential candidate valleys will be precluded and the robot
the obstacle when at the proper distance d,. will not pass through narrow passages.
In summary, it can be concluded that the VFH system
Another important benefit from this method is the elimina- needs a fine-tuned threshold only for the most challenging
tion of the vivacious fluctuations in the steering control (a applications (e.g., travel at high speed and in densely clut-
problem associated with the VFF method). With the averag- tered environments); under less demanding conditions, the
ing effect of the polar histogram and the additional smoothing system performs well even with an imprecisely set threshold.
by ( 5 ) , k, and kf (and consequently 0) vary only mildly One way to optimize performance is to set an adaptive
between sampling intervals. Thus, the VFH method does not threshold from a higher hierarchical level, e.g., as a func-
require a low-pass filter in the steering control loop and is tion of a “global” plan. For example, during normal travel
therefore able to reach much faster to unexpected obstacles. the threshold is set to a very safe, low level. If the global
Similarly, a VFH-controlled robot does not oscillate when plan calls for passing through a narrow passage (e.g., a
traveling in narrow corridors (as is the case with potential doorway), the threshold is temporarily raised while the travel
field methods, under certain circumstances [181). speed is lowered.

C. The Threshold D. Speed Control

As mentioned above, a threshold is used to determine the The robot’s maximum speed V,, can be set at the begin-
candidate valleys. While choosing the proper threshold is a ning of a run. The robot tries to maintain this speed during
critical issue for many sensor-based systems, the perfor- the run unless forced by the VFH algorithm to a lower
mance of the VFH method is largely insensitive to a fine-tuned instantaneous speed V , which is determined in each sampling
threshold. This becomes apparent when considering Fig. 6: interval as follows:
BORENSTEIN AND KOREN:THE VECTOR FIELD HISTOGRAM 285

The smoothed polar obstacle density in the current direc- The maximum travel speed of a VFH-controlled robot is
tion of travel is denoted as h’,. h’, > 0 indicates that an limited by the sampling rate of the ultrasonic sensors and not
obstacle lies ahead of the robot. Large values of h’, mean a by the computation time of the algorithm. In our system, it
large obstacle lies ahead or an obstacle is very close to the takes 160 ms to have all 24 ultrasonic sensors sampled and
robot. Either case is likely to require a drastic change in processed once. We estimate that with our current computing
direction, and a reduction in speed is necessary to allow the hardware (see Section VI) a travel speed of up to 1.5 m/s is
steering wheels to turn into the new direction. This reduction possible if the sampling rate of the sensors could be doubled.
in speed is implemented by the following function: The relationship between sampling time, the robot travel
speed, the signal-to-noise ratio, and the resulting certainty
V’ = V,,, ( 1 - h’f / h,) (6) values is rather complicated and cannot be treated here
where because of space limitations (a through discussion of this
h’f = min(h’,, h,) (7) problem is given in 161 and [71).
and h...
, is an empirically determined constant that causes a V. COMPARISON METHODS
TO EARLIER
sufficient reduction in speed. Note that (7) guarantees V‘ L 0,
During our research in obstacle avoidance for mobile
since h‘f s h,.
robots [2]-[5], we implemented and evaluated some of the
(6) and (7) reduce the speed Of the robot in methods discussed in Section 11. This section the
pation of a steering maneuver, speed can be further reduced
performance of our new VFH method to these earlier meth-
proportionally to the actual steering rate Q:
ods .
V = V‘(1 - Q / Q , , , ) + Vmin A . Comparison to Edge-Detection Methods
where Q,,, is the maximal allowable steering rate for the
The blurry and inaccurate data produced by ultrasonic
mobile robot (in our system ,,a
, = 120°/s). sensors does not provide the sharply defined contours re-
Note that V is prevented from going to zero be setting a quired by edge-detection methods. Consequently, misread-
lower limit for V , namely V 1 V,,,; in our implementation ings or inaccurate range measurements may be interpreted as
Vmin= 4 cm/s. part of an obstacle, thereby distorting the perceived shape of
the obstacle.
E. Limitations and Remedies The VFH method, on the other hand, reacts to clusters of
The VFH method is a local path planner, i.e., it does not range readings. As soon as a range reading has been sam-
attempt to find an optimal path (an optimal path can only be pled, it becomes available to the steering controller (via the
found if complete environmental information is given). Fur- histogram grid) and can influence the path of the vehicle
thermore, a VFH-controlled robot may get “trapped” in immediately. A single range reading will have only minor
dead-end situations (as is the case with other local path influence on the path, while repeated range readings in a
planners). When trapped, mobile robots usually exhibit what confined area (cluster) will cause a more drastic change of
has been called “cyclic behavior,” i.e., going around in direction for the vehicle.
circles or cycling between multiple traps (typical examples The force field method developed by Brooks [8], [9] and
for cyclic behavior are discussed in [5]). While it is possible the similar method developed by Arkin [l], do function in
to devise a set of heuristic rules that would guide the robot experimental real-time systems, using actual sensory data [8],
our of trap situations and resolve cyclic behavior [5], the [9]. However, these methods are somewhat oversimplified,
resulting path is still not optimal. since a threshold determines if an object is at a safe distance
To resolve these problems, we have introduces a path or too close. In the latter case, and because of the binary
monitor that works as follows: If the robot diverts from the character of the threshold, the robot must stop and rotate
target direction khrg (see Fig. 9) the path monitor records away from the object before resuming motion. An additional
this as either left (as is the case in Fig. 9) or right diversion shortcoming of these methods is their susceptibility to mis-
mode. Subsequently, when looking for the near border of the readings (due to ultrasonic noise, crosstalk, etc.) since they
closest candidate valley, k, (see Section IV-B), the VFH take into account only one set of range readings (one reading
algorithm searches to the left or right of ktarg,according to from each ultrasonic sensor). Consequently, misreadings and
the original diversion mode. If k , cannot be found within correct readings (i.e., those produced by actual obstacles)
n = 1 8 0 ° / a = 36 sectors, the path monitor flags a trap have the same weight. Therefore, a single misreading can
situation. Once a certain diversion mode has been set, it is cause the resultant force to exceed the threshold level and
only cleared if the robot faces again into the target direction. “scare” the robot away from a possible safe, free path. Our
When a trap situation is flagged, the robot slows down method, on the other hand, also takes into account past
(and may come to a complete halt), while the VFH algorithm measurements by means of the histogram grid, thereby in-
is temporarily suspended. A global path planner (GPP) creasing the weight of recurring measurements, while mini-
algorithm is then invoked to plan a new path based on the mizing the weight of randomly occurring misreadings. In
available information in the histogram grid [29]. The new addition, the smoothing function [see (5)] reduces the weight
path comprises a set of via points that are then presented as of false readings. Thus, the VFH method results in much
intermediate target locations to the VFH algorithm. more robust and error-resistant control. An additional advan-
 286 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. I, NO. 3, JUNE 1991

the histogram grid may serve for map building purposes and
for the global path planner (see Section IV-E). A large
histogram grid, however, is not necessary for our algorithm
to work properly. It is the short-term effect of the histogram
grid that is important for the VFH algorithm. As explained in
Section IV-A, only cells within the active window influence
the VFH computations. Since the active window travels with
the robot, cells are only briefly inside the window and have
thus only a temporary (short term) effect. Also, since the
ultrasonic sensors are limited to only 2-m look-ahead (about
the size of the active window), only cells inside the window
are updated with sensor information. Therefore, the VFH
algorithm would work equally well if all information was lost
from cells that are outside of the active window. Through the
concept of the active window, the histogram grid becomes a
type of “short-term memory,” where readings are retained
briefly (while the active window sweeps through the area) to
enhance the accuracy by accumulating multiple sensor read-
ings. In a way, this process is similar to the short-term
memory associated with human hearing. Without this mecha-
nism, people would hear but not necessarily comprehend all
speech.
Fig. 10. CARMEL, The University of Michigan’s Mobile Robot. VI. EXPEFUMENTAL
RESULTS
We implemented and tested the VFH method on our
tage of the VFH method is the permanent map information mobile robot, CARMEL (Computer-Aided Robotics for
contained in the histogram grid after a run. Brooks’ and Maintenance, Emergency, and Life support). CARMEL is
Arkin’s methods, on the other hand, do not produce a based on a commercially available mobile platform [12], as
permanent record. seen in Fig. 10. This platform has a maximum travel speed of
A critical discussion of both simulated and experimental V,,, = 0.78 m/s, a maximum steering rate of Q = 120”/s,
potential field methods is given in [18]. Also, based on a and weighs (in its current configuration) about 125 kg. The
rigorous mathematical analysis, [ 181 discusses six inherent platform has a hexagonal structure and a unique three-wheel
shortcomings of potential field methods. drive (synchro-drive) that permits omnidirectional steering.
A 2-80 on-board computer serves as the low-level controller
B. Reflexive versus Reactive Control of the vehicle. Two computers were added: a PC-compatible
On the more abstract level, researchers are beginning to single-board computer to control the sensors, and a 20-Mhz
distinguish between tow fundamentally different approaches 80386-based AT-compatible that runs the VFH algorithm.
to mobile robot obstacle avoidance. The “conventional” CARMEL is equipped with a ring of 24 ultrasonic sensors
approach, reactive control, is based on the traditional artifi- [25]. The sensor ring has a diameter of 0.8 m, and objects
cial intelligence model of human cognition. Reactive control must be at least 0.27 m away from the sensors to be detected.
algorithms reason about the robot’s perception (sensor data) Therefore, the theoretical minimum width for safe travel in a
while building elaborate world models (memory) and subse- +
passageway is Wmin= 0.8 2 * 0.27 = 1.34 m.
quently plan the robot’s actions. This approach requires large In extensive tests, we ran the VFH-controlled CARMEL
amounts of computation and decision making, resulting in a through difficult obstacle courses. The obstacles were un-
relatively slow response of the system. Reflexive control marked, commonplace objects such as chairs, partitions, and
(with Brooks as one of its foremost proponents), eliminates bookshelves. In most experiments, CARMEL ran at its maxi-
cognition altogether. In reflexive control there is no planning mum speed V,, = 0.78 m/s. This speed was only reduced
and reasoning; nor are there world models. Simple reflexes when an obstacle was approached head-on (see discussion of
tie actions to perceptions, resulting in faster response to speed control in Section IV-D).
outside stimuli. Fig. 11 shows the histogram grid after a run through a
At first glance it may seem that our VFH method is a particularly challenging obstacle course of 3 /Gin-diameter
typical example of reactive control, considering the his- vertical poles spaced at a distance of about 1.4 m from each
togram grid world model and even a second model, the polar other. The actual location of the rods is indicated by plus (+)
histogram. However, some distinctions should be made. Our symbols in Fig. 11. It should be noted that none of the
world model, the histogram grid, has two different functional obstacle locations were known to the robot in advance: the
properties, namely a short-term effect and a long-term effect. CV clusters in Fig. 11 gradually appeared on the operator’s
The long-term effect is provided by the whole histogram screen while CARMEL was moving.
grid, as described in Section III. The information stored in To test the performance limits of our system, we per-

__
BORENSTEIN AND KOREN: THE VECTOR FIELD HISTOGRAM 287

histogram (VFH) method, has been developed and success-

....... ..
fully tested on our experimental mobile robot CARMEL. The
VFH algorithm is computationally efficient, very robust, and
t o size 1.h insensitive to misreadings, and it allows continuous and fast
motion of, the mobile robot without stopping for obstacles.
The VFH-controlled mobile robot traverses very densely
cluttered obstacle courses at high average speeds and is able
to pass through narrow openings (e.g., doorways) or negoti-

Robow
path
ate narrow corridors without oscillations.
The VFH method is based on the following principles:

.... .._
.. /
Start
Y

.-. ..... .-.. I;

I ShOG as
CV.1-3
CV.4-6
CV=l-9
1
A two-dimensional Cartesian histogram grid is continu-
ously updated in real-time with range data sampled by
the on-board range sensors.
I
.. J The histogram grid is reduced to a one-dimensional
Wall polar histogram that is constructed around the momen-
Fig. 1 1 . Histogram grid representation of a run through a field of densely tary location of the robot. The polar histogram is the
spaced, thin vertical poles. The average speed is this run was Vavrg = 0.58
m/s.
most significant distinction between the VFF and the
VFH methods as it allows a spatial interpretation (called
polar obstacle density) of the robot’s instantaneous
formed a variety of experiments with other slender obstacles. environment.
For example, 1/2-in-diameter poles were still detected, but Consecutive sectors with a polar obstacle density below
not reliable so when approached at CARMEL’s maximum threshold are called “candidate valleys. The candidate
”

speed. Unreliable detection would also result when 1 in x 1 valley closest to the target direction is selected for
in square rods were used. Larger objects, such as 2-in-diame- further processing.
ter cylinders, square-shaped cardboard boxes, furniture, and The direction of the center of the selected candidate
even slowly walking people are reliable detected and avoided. direction is determined and the steering of the robot is
These obstacles are easier to detect than the 3/4-in poles in aligned with that direction.
the experiment described here. The speed of the robot is reduced when approaching
Each blob in Fig. 11 represents one cell in the histogram obstacles head-on.
grid. In our current implementation, certainty values (CV’s)
range from 0 to 15 and are indicated in Fig. 11 by blobs of The characteristic behavior of a VFH-controlled mobile
varying sizes. CV = 0 means no sensor reading has been robot is best described as follows: The response of the
projected onto the cell during the run (i.e., no blob), while vehicle is dependent on the likelihood for the existence of
CV’s > 0 indicate the increasing confidence in the existence an obstacle. In the histogram grid, this likelihood is ex-
of an object at that location. CARMEL traversed this obsta- pressed in terms of size and certainty values of a cluster. This
cle course at an average speed of 0.58 m/s without stopping information is transformed into the height and width of an
for obstacles. Note that this is a typical experimental run, and elevation in the polar histogram. The strength of the VFH
similar performance has been routinely obtained in countless method lies thus in its ability to maintain a statistical obstacle
experiments and demonstrations, using different kinds of representation at both the histogram grid level as well as at
obstacles at random layouts. the intermediate data level, the polar histogram. This feature
An indication of the real-time performance of the VFH makes the VFH method especially suited to the accommoda-
algorithm is the sampling time T (i.e., the rate at which the tion of inaccurate sensor data, such as that produced by
steer and speed commands for the low-level controller are ultrasonic sensors, as well as sensor fusion.
issued). The following steps occur during T :
1) Obtain sonar information from the sensor controller. REFERENCES
2) Update the histogram grid. [ l ] R. C. Arkin, “Motor schema-based mobile robot navigation,” Int.
J . Robotics Res., pp. 92-112, Aug. 1989.
3) Create the polar histogram.
[2] J . Borenstein and Y . Koren, “A mobile platform for nursing robots,”
4) Determine the free sector and steering direction. IEEE Trans. Industrial Electron., vol. 32, no. 2, pp.158-165,
5) Calculate the speed command. 1985.
6 ) Comunicate with the low-level motion controller (send [3] -, “Motion control analysis of a mobile robot,” Trans. ASME,
J . Dynamics, Measurement Control, vol. 109, no. 2, pp. 73-79,
speed and steer commands and receive position update). 1987.
on an Intel 80386-based PC-compatible running t41 -, “Obstacle avoidance With ultrasonic sensors,” IEEE J .
Robotics Automat., vol. R A 4 , no. 2 , pp. 213-218, 1988.
at 20 MHz, T = 27 ms. 151 -, “Real-time obstacle avoidance for fast mobile robots,” IEEE
Trans. Systems Man Cybern., vol. 19, no. 5, pp. 1179-1187,
VU. CONCLUSIONS Sept./Oct. 1989.
[6] -, “Histogramic in-motion mapping for mobile robot obstacle
This paper presents a new obstacle-avoidance method for avoidance,” IEEE Trans. Robotics Automat., vol. 7, to be pub-
fast-running vehicles. This approach, called the vector field lished, Aug. 1991.
288 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 7, NO. 3, JUNE 1991

171 -, “Real-time map-building for fast mobile robot obstacle avoid- [26] U. Raschke and J . Borenstein, “A comparisson of grid-type map-
ance,” presented at SPIE Symp. on Advances in Intelligent Systems, building techniques by index of performance,” in Proc. IEEE Int.
Mobile Robots V, Boston, MA, Nov. 4-9, 1990. Conf. Robotics Automat. (Cincinnati, OH, May 13- 18, 1990).
181 R. A. Brooks, “A robust layered control system for a mobile robot,” [27] C . F. Thorpe, “Path relaxation: Path planning for a mobile robot,”
IEEE J . Robotics Automat., vol. RA-2, no. 1, pp. 14-23, 1986. Carnegie-Mellon Univ. The Robotics Institute, Mobile Robots Lab.,
191 R. A. Brooks and J . H. Connell, “Asynchronous distributed control Autonomous Mobile Robots, Annual Rep., pp. 39-42, 1985.
system for a mobile robot,” Proc. SPIE, Mobile Robots, vol. 727, [28] C. R. Weisbin, G. de Saussure, and D. Kammer, “SELF-CON-
pp. 77-84, 1987. TROLLED. A real-time expert system for an autonomous mobile
3 . L. Crowley, “Dynamic world modeling for an intelligent mobile robot,” Computers Mechanic. Eng., pp. 12-19, Sept. 1986.
robot,” in Proc. IEEE 7th Int. Conf. Pattern Recognition [29] Y. Zhao, S. L. BeMent, and J . Borenstein, “Dynamic path planning
(Montreal, Canada, July 30-Aug. 2, 1984), pp. 207-210. for mobile robot real-time navigation,” presented at 12th IASTED
-, “World modeling and position estimation for a mobile robot Int. Symp. Robotics and Manufacturing, Santa Barbara, CA, Nov.
using ultrasonic ranging,” in Proc. IEEE Int. Conf. Robotics 13-15. 1989.
Automat. (Scottsdale, AZ, May 14-19, 1989), pp. 674-680.
K2A Mobile Platform, Cybernation, Roanoke, VA, 1987.
A. Elfes, “Sonar-based real-world mapping and navigation,” IEEE
J . Robotics Automat., vol. RA-3, no. 3, pp. 249-265, 1987.
A. M. Flynn, “Combining sonar and infrared sensors for mobile
robot navigation,” Int. J . Robotics Res., vol. 7, no. 6, pp. 5-14,
Dec. 1988. Johann Borenstein (M’88) received the B.Sc.,
1151 W. S. Newman and N. Hogan, “High speed robot control and M.Sc., and D.Sc. degrees in mechanical engineer-
obstacle avoidance using dynamic potential functions,” in Proc. ing in 1981, 1983, and 1987, respectively, from
IEEE Int. Conf. Robotics Automat. (Raleigh, NC, Mar. 31-Apr. the Technion-Israel Institute of Technology,
3, 1987), pp. 14-24. Haifa, Israel.
1161 0. Khatib, “Real-time obstacle avoidance for manipulators and mo- Since 1987, he has been with the Robotics Sys-
bile robots,” in Proc. IEEE Int. Conf. Robotics Automat. (St. tems Division at the University of Michigan, Ann
Louis, MO, Mar. 25-28, 1985), pp. 500-505. Arbor, where he is currently an Assistant Research
Y. Koren and J . Borenstein, “Analysis of control methods for mobile Scientist and Head of the MEAM Mobile Robotics
robot obstacle avoidance,” in Proc. IEEE Int. Workshop Intelli- Laboratory. His research interest include mobile
gent Motion Control (Istanbul, Turkey, Aug. 20-22, 1990), pp. robot navigation, obstacle avoidance, path plan-
457-462.
- ning, real-time control, sensors for robotic applications, multisensor integra-
1181 -, “Critical analysis of potential field methods for mobile robot tion, and computer interfacing and integration
obstacle avoidance,” IEEE Trans. Robotics Automat., submitted
for publication.
1191 B. H. Krogh, “A generalized potential field approach to obstacle
avoidance control,” presented at Int. Robotics Res. Conf., Bethlehem
PA, Aug. 1984.
B. H. Krogh and C. E. Thorpe, “Integrated path planning and
dynamic steering control for autonomous vehicles,” in Proc. IEEE Yoram Koren (M’76-SM’88) has 25 years of
Int. Conf. Robotics Automat. (San Francisco, CA, Apr. 7-10, research, teaching, and consulting experience in
the automated manufacturing field. At present he is
1986), pp. 1664-1669.
R. Kuc and B. Barshan, “Navigating vehicles through an unstructured a Professor in the Department of Mechanical Engi-
environment with sonar,” in Proc. IEEE Int. Conf. Robotics neering and Applied Mechanics at the University
Automat. (Scottsdale, AZ, May 14-19, 1989), pp. 1422-1426. of Michigan, Ann Arbor. He has published over
V. Lumelsky and T. Skewis, “A paradigm for incorporating vision in 100 papers and three books on machine tool con-
the robot navigation function,” in Proc. IEEE Conf. Robotics trol, robotics, sensing methods, and modeling of
Automat. (Philadelphia, PA, Apr. 25, 1988), pp. 734-739. processes. His book Computer Control f o r Man-
H. P. Moravec and A. Elfes, “High resolution maps from wide angle ufucturing Systems (McGraw-Hill, 1983) is used
sonar,” in Proc. IEEE Conf. Robotics Automat. (Washington, as a textbook at maior universities and received the
1984 Textbook Award for the Society of Manufacturing Engineering (SME).
DC), 1985, pp. 116-121.
His book, Robotics f o r Engineers (McGraw-Hill, 1985), was translated
~ 4 1H. P. Moravec, “Sensor fusion in certainty grids for mobile robots,” into Japanese and French and is used by engineers throughout the world.
AI Mag., pp. 61-74, Summer 1988.
Prof. Koren is a Fellow of the ASME, an Active Member of CIRP, and a
1251 POLAROID Corporation, Ultrasonic Components Group, Cambridge, Fellow of SME/Robotics-International.
MA, 1990.