Content uploaded by Tomas Nordfjell
Author content
All content in this area was uploaded by Tomas Nordfjell
Content may be subject to copyright.
1
AU T O N O M O U S F O R E S T
M AC H I N E S – PA S T, P R E S E N T
A N D F U T U R E
Thomas Hellström
1
, Pär Lärkeryd
2
, Tomas Nordfjell
3
and Ola Ringdahl
4
(1) Senior lecturer, Dept. of Computing Science, Umeå University, SE-901 87 Umeå,
Sweden. thomash@cs.umu.se
(2) CEO, Indexator AB, Lidvägen 2, SE-922 31 Vindeln, Sweden. par.larkeryd@indexator.se
(3) Professor, Dept. of Forest Resource Management, Swedish University of Agricultural
Sciences, SE-901 83 Umeå, Sweden. tomas.nordfjell@srh.slu.se
(4) Research engineer, Dept. of Computing Science, Umeå University, SE-901 87 Umeå,
Sweden. ringdahl@cs.umu.se
UMINF 08.06
ISSN-0348-0542
April 21, 2008
2
ABSTRACT
The feasibility of using autonomous forest vehicles (which can be regarded as logical
developments in the ongoing automation of forest machines), the systems that could be
applied in them, their potential advantages and their limitations (in the foreseeable future) are
considered here. The aims were to analyse: (1) the factors influencing the degree of
automation in logging; (2) the technical principles that can be applied to autonomous forest
machines, and (3) the feasibility of developing an autonomous path-tracking forest vehicle. A
class of such vehicles that are believed to have considerable commercial potential is
autonomous wood shuttles (forwarders). The degree of automation is influenced by increased
productivity, the machine operator as a bottle-neck, cost reduction, and environmental
aspects. Technical principles that can be applied to autonomous forest vehicles are satellite
navigation, laser odometry, wheel odometry, laser scanner and radar. The presented system
has demonstrated both possibilities and difficulties associated with autonomous forest
machines. It is in a field study shown that it is quite possible for them to learn and track a path
previously demonstrated by an operator with an accuracy of 0.1m on flat ground. A new path-
tracking algorithm has been developed to reduce deviations by utilizing the driver’s steering
commands.
Keywords: Forest technology, obstacle detection, path-tracking, robotic, system architecture.
3
1. INTRODUCTION
The aims of this study were to analyse: (1) the factors influencing the degree of automation in
logging; (2) the technical principles that can be applied to autonomous forest machines, and
(3) the feasibility of developing an autonomous path-tracking forest vehicle.
The first and second of the aims listed above were addressed by literature studies and
theoretical analysis, while the third aim was addressed in a field study.
The paper is organized as follows: Section 2 gives a historical background to automation in
forestry. Section 3 analyzes the major driving forces behind the urge to develop autonomous
forest machines while Section 4 summarizes the basic requirements for such systems. In
Section 5, a number of basic system design scenarios for semi-autonomous vehicles are
described. In Section 6, the experiences from a pilot project aiming at developing an
autonomous forest vehicle are discussed. The results from the field study are presented in
Section 7 and conclusions regarding state-of-the-art and future development in the area are
given in Section 8.
2. HISTORICAL BACKGROUND
For a long time forestry operations were largely performed manually, but the scope for further
improvements to the tools used, such as saws, axes and horse-drawn sledges, had become
very limited by around the year 1900 (Ekman 1908, Brown 1949, Sundberg 1978, Silversides
1997). However, the mechanisation of forestry started much later than the mechanisation of
agriculture (Sundberg 1978, Silversides 1997), with the introduction of the chain-saw for
harvesting and tractors for extraction. Nevertheless, there were some early examples of
technical forestry innovations. The first chain-saws light enough to be handled by just two
persons were developed in the USA and Sweden in the years 1916-17 (Sundberg 1978,
Silversides 1997). In USA, a single factory was producing chain-saws in 1938, six in 1942
and 30 in 1949 (Silversides 1997). The definitive breakthrough for chain-saws occurred in
around 1950, when they became sufficiently light to be handled by a single operator
(Sundberg 1978, Drushka & Konttinen 1997, Silversides 1997). In 1952, only 20% of the
pulp wood produced in eastern Canada was harvested with chain-saws, but by 1960 the
proportion had risen to almost 100% (Silversides 1997). Trends in the European countries
were similar, and a number of chain-saw factories were built in Sweden, and Germany.
However, chain-saws were still being used by fulltime professional operators at thinning
operations in Scandinavia until the beginning of the 1990s (Lidén 1995).
Many attempts were made in North America from the 1920s onwards to use tractors in
forestry (Brown 1949, Silversides 1997). An important step in mechanisation was that from
1925 tractors could be equipped with winches (Brown 1949). More than 8000 machines were
in use in Canadian forestry in 1950; three times more than in 1945 (Silversides 1997). Many
tracked vehicles for extraction were invented in the Soviet Union around 1950 and in North
America during the 1950s (Andersson 2004), were a skidder with articulated steering was
invented in 1959 (Silversides 1997). In addition, several machines for mechanised tree felling,
loading and extraction were invented during the 1950s, as well as machines for debranching
and cutting logs at road-side (Silversides 1997).
The large-scale mechanisation of extraction and harvesting operations occurred somewhat
later in the Scandinavian countries than in North America, although some early examples of
Scandinavian tractors for extraction were produced by the beginning of the 1930s (Drushka &
Konttinen 1997). Total global sales of forest machines in 1966 amounted to ca. 400 CTL
machines (generally used in Scandinavian forestry) and more than 5500 machines for stem- or
tree-cutting methods (generally used in North American or Soviet Union forestry) (Drushka &
4
Konttinen 1997). In 1957 the first Swedish tailor-made forest tractor was introduced, and
manufactured in large numbers (Östberg 1990). The introduction in 1959 of the first hydraulic
grapple loaders, mounted on tractors, was a highly important development for Scandinavian
forestry (Malmberg 1988). A forwarder with articulated steering, inspired by the American
“Blue ox” skidder and invented in 1962, was the next important Swedish breakthrough (Staaf
1988). Nevertheless, in the year 1960, the horse was still used for extracting more than 80%
of the total volume of logs extracted in Sweden, but by 1970 mechanised extraction was
totally dominant with 95% of the volume (Andersson 2004). Machines for debranching and
log cutting were introduced in the years 1966-67 (Nordansjö 1988) and machines for tree-
felling in 1972 (Östberg 1990). The whole CTL-harvesting operation was then mechanised.
Harvesters, which fell trees, debranch and cut logs, were invented in Sweden and Finland in
the years 1972-73 (Drushka & Konttinen 1997). In 1982 in professional forestry in Sweden,
51% of the volumes from final fellings were cut with a two-machine system (one for tree
felling and one for debranching and log cutting), 21% of the volume was cut with a harvester
and 24% was cut motor manually (using a chain-saw). The corresponding values five years
later were 25, 44 and 15%, respectively (Fryk et al. 1991).
The basic principles for CTL forest machines have remained the same since the 1990s.
Developments since then have been mainly focused on raising productivity, reducing costs
and optimising the division of the trees into logs.
The mechanisation of logging can be divided into six phases according to Silversides (1997),
in which the implements used were predominantly: hand tools and draught animals; various
combinations of hand tools; motor-manual tools; manually operated machines; machines that
automatically performed some repetitive work elements and machines that use feedback from
the process to control the next work element (e.g. a harvester with a bucking computer).
However, a further phase of mechanisation may occur when machines with no operators that
can work autonomously are developed (Gellerstedt et al. 1996). The machine presented by
Golob (1981) that required no operator for tree felling, debranching and laying stems in piles
can be seen as an early conceptual example of such a machine. The forces driving
mechanisation are a lack of workers, the aspiration to continue forestry operations year-round
and for more hours per day, and the desires to reduce costs, the amounts of hard physical
work involved and the lead-times between logging and industrial processing (Sundberg 1978,
Silversides 1997).
3. WHY AUTONOMOUS VEHICLES IN FORESTRY?
Autonomous vehicles have been considered, and tested, in various research and development
projects, in efforts to make forestry operations more efficient. The main identified benefits are
as follows.
Increases in productivity
Increases in productivity per unit time are perhaps the most important potential benefits of
automation: timber production will amount to 60 000 m
3
per year if a forest machine produces
20 m
3
per hour over an annual work-time of 3000 hours (with two drivers working in
shifts).To increase the annual work-time to more than 3000 hours using the same machines
more drivers would be required, and complex logistics with huge numbers of machine
movements between harvesting sites. However, there is another way to achieve the same
annual production levels. A year has 8760 hours, and if a machine can work autonomously it
might be possible for it to work efficiently over perhaps 6000 hours per year. The same
annual production level (60 000 m
3
) could then be delivered by a machine producing only 10
5
m
3
per hour. Furthermore, the cost per hour of an autonomous vehicle is not directly
influenced by salary costs, because it has no driver.
Elimination of the machine operator as a bottle-neck
Crane-speeds are currently limited essentially by the speeds their operators can handle, which
equate to average productivity levels of ca. 20 m
3
per hour, while the load capacity of
forwarders is currently limited solely by the weight the forest ground can support. A harvester
can fell, delimb and buck 60 – 100 trees per hour, and a large forwarder can carry a load of 18
tonnes. The machines are fast, but require an operator who constantly manoeuvres its crane
and harvester head and distinguishes differences in log quality classes both between and
within trees. This makes the work environment stressful, since many decisions must be taken
at high pace, and the operator often becomes a production bottle-neck. The operator can also
be the limiting factor for rates of loading and unloading forwarders. A possible way to
increase productivity is therefore to raise the level of automation and, if possible, use
autonomous forest machines.
Cost reductions
The salary of operators generally amounts to 30-40% of the hourly cost of a forest machine.
Thus, there would be substantial economic advantages if a machine could work
autonomously, or an operator could handle more than one machine at the same time.
Environmental aspects
The size and load capacities of forwarders have increased to raise productivity, and the risks
of damage to the ground have increased accordingly. Harvesters have also become larger.
This damage could be reduced, while maintaining overall productivity levels, by using an
autonomous harvesting system working 6000 productive hours per year but at only half the
hourly productivity rate of current forest machines. Thus, less advanced basic machines could
be used, e.g. harvesters processing 30-50 trees per hour with much slower crane movements,
and consequent reductions in mechanical stress, and forwarders with gentler crane
movements, slower driving speeds, and smaller loads. Generally, lower speeds require less
engine power. Thus, an autonomous machine can be lighter since its speeds and loads can be
lower, and the removal of the driver cabin alone can reduce the mass by ca. a tonne. It is also
easier to optimise the weight distribution of a forest machine without a large cabin. For
instance, it might be possible to place the crane of a forwarder in the middle of the machine,
and piles of wood on both its front and rear parts. With such changes a forwarder should be
able to carry a load of at least the same mass as itself. A 9-10 tonne forwarder would then be
able to carry a load of 10 tonnes, and the total mass would still not exceed 19-20 tonnes. Such
a machine would place considerably less stress on the forest ground and tires than current
forest machines. A lighter machine, with a higher load/mass ratio, would also consume less
fuel and generate less emission per m
3
handled.
4. WHAT IS REQUIRED FOR AUTONOMOUS WORK SYSTEMS IN
FORESTRY?
For the foreseeable future human beings will play a central role in the manoeuvring of forest
machines in harvesting operations, because so many complex factors need to be considered in
the work done by forest machines that full automation would be extremely difficult. The
issues that must be addressed include the following:
6
What trees should be harvested, and what are their optimal bucking lengths and
assortments, according to both economic and environmental considerations?
Where should the harvester be positioned, and what routes should be taken? These
decisions are based on ground conditions, obstacles in the terrain, judgements regarding
suitable driving routes in relation to the reach of the crane, the position and size of the
trees and the stability of the machine, which depends on the inclination of the ground,
slant of the tree, wind affecting the tree and environmental considerations.
In what direction should a tree be felled? The tree-felling has to be done without risks to
human beings, the machine itself or surrounding objects, for example power transmission
lines. The decision is based on the slant of the tree, the wind, the risk of damaging
remaining trees and the planned positions of small piles of saw-logs and pulpwood for
ease of extraction.
The transportation of wood out of the forest to the road-side requires an ability to
navigate, and thus the ability to follow a specific route. Navigation demands real time
knowledge with high precision about the surroundings as well as the position of the
machine in each moment.
The complex work elements and decisions described above have to be automated before an
autonomous forest machine can be constructed. This means that extensive research has to be
conducted within the three following main areas.
Support for decision-making
The driver of a harvester has to take many decisions (regarding, inter alia, where the machine
should be positioned, where small piles of logs should be placed and what trees should be
harvested), and systems that could facilitate the decision-making could be extremely useful.
The most complex decisions concern selective thinning, regarding which trees should be
harvested depending on the species and quality of the individual trees, and the properties of
the surrounding stand (density, tree sizes, species composition, etc.) (cf. Vestlund 2005). A
semi-autonomous way to solve this problem was presented by Kurabyashi & Asaman (2001),
in which trees selected for retention were identified by discs attached to them. A decision
support system for selective cleaning and thinning was developed by Vestlund et al. 2005, but
sensors that are more capable of detecting the trees and measuring the stand characteristics
that the algorithm relies on need to be developed before it can be fully implemented (Vestlund
2005).
Extraction of logs to the road-side does not require many difficult decisions to be taken. The
route is already roughly known because the harvester has already driven from the road-side to
the harvesting site, and manoeuvred at the site when harvesting trees. The harvester has also
placed small piles of saw-logs and pulp-wood beside the route it took within the harvesting
site. Therefore, before the extraction begins it should be possible to acquire information about
both routes and the positions of wood piles.
Automation of work elements
A number of repetitive work elements can be partly or fully automated, e.g. crane movements
during loading, unloading and sorting of assortments, positioning of the grapple or harvester
head, and placement of the crane in a suitable position when driving the machine.
Autonomous navigation in terrain
All autonomous work tasks in a forest require an ability to navigate. An early attempt to
develop an autonomous navigation system was presented by Kourtz (1996). To be successful
7
the vehicle must know were it is at all times, and how it should manoeuvre to follow the
decided route. Systems for detecting obstacles, human beings, animals, other machines and
buildings are also needed, as well as advanced transmission systems to avoid unnecessary
slippage.
5. POSSIBLE SCENARIOS
Due to the complexities of the problems involved, the current aim is not to fully automate the
felling process, but rather to develop various types of semi-autonomous systems in which man
is still involved, especially in decision-making. Three examples of possible system designs
are described below. Other possibilities have also been described; see for instance
Hallonborg (2003).
Remote supervision
This approach is based on a human remotely supervising a semi-autonomous system. When
necessary, the operator can take control of the system. Examples of this can be found in the
mining industry, for example LKAB have used unmanned loaders for several years that are
capable of autonomously navigating in mines. In forest applications, a process operator could
supervise such systems from his or her mobile office, and to some degree remotely control a
number of semi-autonomous harvesters and forwarders.
Semi-autonomous harvesters
In this scenario, a manned forwarder remotely controls one or more semi-autonomous
harvesters. An example of this is the prototype system “Besten” [the Beast] (Bergkvist et al.
2006), which consists of an unmanned harvester remotely controlled from a manned
forwarder. Studies indicate that this could reduce the costs of final felling by more than 20
percent (Bergkvist et al. 2006).
Autonomous wood shuttles
This is a system that has not yet been tested in practice, but has been deemed in a study by
Hallonborg (2003) to have the best ability to compete with the current harvester system. In
this scenario, a manned harvester cuts the trees, but the logs are transported to the road by one
or more autonomous shuttles.
8
6. CURRENT RESEARCH TOWARDS DEVELOPING AN
AUTONOMOUS FOREST VEHICLE
During the last few years we have been addressing various problems that need to be solved
before viable autonomous forest machines can be constructed, and exploring systems that
could be applied in them, at the IFOR-centre, Umeå University. Algorithms for controlling
hydraulic cranes have been developed, which it is hoped will lead to semi-autonomous crane
functions for next-generation cranes of forest machines. The main aim for another project is to
establish a solid research platform for investigating in detail the feasibility, optimal features,
limitations and prospective systems of autonomous wood shuttles such as those described in
the previous section (Hellström et al. 2006, Ringdahl 2007). The project is based on a solution
in which an operator demonstrates a path by driving the vehicle manually along it once, either
by remote control or from the cabin (if still present). The computer in the shuttle records both
the vehicle’s pose and the operator’s steering commands. Thus, the system can learn the
adjustments to controls required to drive along the demonstrated path. However, due to
inaccuracies in signals from the position sensors in combination with unevenness of the
ground, deviations from the path can occur. Consequently, the problems involved in
autonomously tracking a demonstrated path are not easy to resolve, and interaction with the
environment is required for satisfactory solutions. The scenario in which an operator initially
demonstrates the path is of course a simplification and is not intended to be a final solution for
use in a commercial product. An alternative to manual demonstration of the path could be to
use information from a map. However, even in this case the ability to follow a predefined path
is essential. The chosen scenario should therefore also provide a suitable test platform for
further research relevant to these issues.
Figure 1 shows an overview of the system developed in the project to date. The computing
power is split between two computers connected by a Wireless Local Area Network (WLAN).
The primary computer is mounted in the autonomous shuttle and is responsible for hardware
interfacing and communication with the vehicle. It also contains low-level routines for
controlling velocity and steering angle, as well as processing sensor information. The
secondary computer is used by the operator to initiate and supervise the autonomous shuttle’s
operation.
The main technical challenges that need to be resolved to enable autonomous forest vehicles
are more refined techniques for positioning and obstacle detection. Solutions for these
problems are needed for the main task of path tracking. A few possible approaches for these
areas are presented below.
9
Actuators
Steering force
Throttle pedal
Brake
Sensors
For localization:
GPS/GLONASS
Gyro
For obstacle detection:
24 GHz radar
Laser scanner
Engine rpm
Steering angle
Mobile computer
Win XP Java
- Hardware interface
to sensors and actuators
- Low-level control loop
- Low-level data analysis and
occupancy grid
- Communication
Remote computer
Win XP Matlab
- User interface
- Path learning
- High-level control loop
for path tracking
- Data analysis routines
- Communication
WLAN
communication
DGPS base station
Figure 1. System overview of prototype machine. Two computers share the computing tasks;
one handles low-level tasks such as processing sensor information, while the other handles
higher level tasks, including path tracking, learning new paths, and user interfaces. The forest
machine shown is a concept of what a future autonomous shuttle could look like.
Positioning
For positioning, an accuracy of ±1 meter is regarded as both realistic and essential for safe
operation in a forest environment. This is the required accuracy for estimates of the real
position of the vehicle, so the positional information from the sensors must not deviate more
than a couple of decimetres from true values to ensure safe control of a large forest machine.
The following sensor techniques have been tested in the project:
Satellite navigation - The main sensor used for positioning is an advanced Real-Time
Kinematics Differential GPS (RTK DGPS) from Javad. With clear views of at least four or
five satellites, this system has an accuracy of about 2 centimetres. The receiver is capable of
receiving signals from both the American GPS system and the Russian GLONASS system.
While providing lower accuracy than GPS, GLONASS provides important backup, especially
at the high latitudes (64 degrees north) where the work is being conducted. The vehicle is
equipped with two receivers and two antennas, which also makes it possible to determine the
vehicle’s heading. The system also includes a stationary receiver that compensates for
common sources of errors in the satellite navigation system. This is done using the so-called
differential-GPS technique (DGPS), which sends correction signals by radio to the mobile
receivers. In the near future, the European Galileo system will hopefully contribute additional
positional information. However, all satellite navigation systems are sensitive to the reception
conditions, and in a forest environment the accuracy can easily deteriorate to 0.5 metres or
worse. In order to navigate even when the satellite navigation system is not providing
sufficient accuracy, techniques for combining several sensors have been developed. Using this
approach the accuracy can be increased sufficiently to autonomously navigate the vehicle for
10
a limited time, for example through a dense part of the forest. The sensing techniques used for
this purpose are briefly described below.
Laser odometry - Algorithms for a positioning technique called laser odometry have been
developed and evaluated in the project. A laser scanner mounted at the front of the vehicle
emits pulsed laser beams in a 180-degeee sweep and measures the distance to any object that
reflects the beam back to the scanner. This results in a picture of the nearest surroundings
consisting of 181 points. More technical details on the laser scanner are given below. To
calculate how the vehicle has moved between the times that two pictures were acquired, they
are compared to each other. In simplified terms, the algorithm that does this translates and
rotates one picture until as many as possible points in the two pictures coincide. This
translation and rotation correspond to the changes in the vehicle’s pose between the two
measurements. To obtain information on the vehicle’s position and heading, all movement
changes are summed. A drawback associated with this method is that many small errors
accumulate, aggregating to total errors that increase with time. This drawback also affects
many other localization techniques. Consequently, sufficient accuracy can only be maintained
for short periods of time. Poor accuracy has been obtained in initial tests of laser odometry in
a forest environment, but further tests are needed to determine how well this technique could
work in practice
.
Wheel odometry - A vehicle’s movements can be calculated from measurements of its
velocity and steering angles. A problem with this technique is that the vehicle’s wheels can
slip considerably, resulting in reduced accuracy. Another problem, related to the articulated
joint design, is that the motions of the front and rear parts of the vehicle relative to the ground
are uncertain. When the steering angle changes the two parts move differently, depending on
factors such as weight distribution and ground conditions, leading to the introduction of small,
cumulative errors in each step of the calculations when the machine is changing heading, and
thus drift similar to the drift affecting laser odometry, as described above. To improve wheel
odometry, neural networks have successfully been used in the project (Ringdahl 2007). By
training a neural network with velocity and steering angle data from the forest machine and
using the exact positions obtained from the satellite navigation system, neural networks are
able to learn the mathematical relationships between velocities/steering angles and changes in
pose. The trained network can then be used as position sensors that are considerably less
subject to drift than ordinary wheel odometry systems. In trials in a forest environment the
average time before the drift exceeded 1 metre was 13 seconds when a neural network was
used, compared to 3 seconds when uncompensated wheel odometry was used.
Gyro - Normally the satellite navigation system is used to determine the vehicle’s heading
and, as for positional data, GPS signals provide highly accurate directional accuracy under
normal circumstances. However, if the GPS antennas do not have clear views of sufficient
satellites, a gyro with an internal compass is used instead (although the compass is not used
since the magnetic material in a forest machine substantially affects the magnetic fields in and
around it). The gyro we have been using in tests to date is an AHRS400CC from Crossbow
Technology, based on MEMS-technology (Micro Electro-Mechanical Sensors). Instead of
moving parts, vibrating ceramic plates are used to sense the angular rate. By multiplying the
mean angular rate reported by the gyro in a given sampling period by its duration we can
obtain a good estimate of the vehicle’s change in heading during the sampling period. The
vehicle’s current heading is calculated by summing these changes. Like the values provided
by odometry-based methods for determining position, the values obtained here are subject to
drift over time. However, the gyro system is much more accurate than systems based on either
11
wheel odometry or laser odometry. Estimates of heading from the gyro can be used with
sufficient accuracy (±5 degrees) for movements lasting about 50 seconds.
Obstacle detection
Detecting obstacles in forest environments is extremely demanding and raises problems that
have not yet been satisfactorily resolved. In particular it is difficult to determine if a detected
object is a real obstacle, like a stone, or a traversable object like a bush. Rough terrain and
“negative obstacles” (e.g. holes in the ground) present further challenges.
In the project, information from a laser scanner and a radar sensor is being used to create and
update local maps of the environment. The resulting occupancy grid includes probabilities of
obstacles in a fine-meshed grid and is created by weighing and integrating a large number of
sensor readings
.
Laser scanner - An LMS 221 laser scanner supplied by SICK, which operates by measuring
the time of flight of laser light pulses, is used to detect obstacles in front of the forest
machine. A pulsed laser beam is emitted from the scanner, reflected back if it meets an
obstacle, and is then registered by the scanner’s receiver. The time between transmission and
reception of the pulse is directly proportional to the distance to the obstacle. The maximum
range at which the scanner can detect an obstacle depends on the reflectivity of the object; the
more reflective the object, the further away the scanner can detect it. A tree can be detected at
about 60 metres. To reduce the scanner’s sensitivity to rain or snow, the scanner takes several
consecutive scans in thea same direction. By comparing these scans, it is able to determine
whether there is a real obstacle at a spot from which it detected a signal or if the signal was
due to something temporary like a snow flake. The LMS 221 laser scanner has an accuracy of
±35 mm and is able to do a 180 degree scan in 13-53 ms, depending on the angular resolution.
Radar -. An advantage with radar is its lower sensitivity to bad weather, such as fog, rain, or
snow. The radar used in the project is manufactured by TYCO and is primarily used in the car
industry. It has no movable parts, but emits pulses from two antennas with different lobe-
characteristics. By comparing the relative strength of the radar echoes the distance and
bearing to several targets can be computed. The resolution is about 15 centimetres and the
radar unit can detect obstacles up to 30 metres away.
Path-tracking
The vehicle has to be able to answer autonomously the following three questions in order to
follow a route: Where am I? Where should I go? How do I get there? The first question is
answered with the help of positioning sensors and systems, an important component of which
is the satellite navigation system. Where the vehicle should go, and how to get there are
mainly defined by the operator who demonstrates the desired route. Both the final destination
and the route are defined in this way. Algorithms for path-tracking are used to provide the
ability to follow the learnt route. One of the simplest algorithms is named Follow the Carrot,
which means that the vehicle steers straight towards a point further along the route, with no
concern about the appearance of the route before that point (Barton, 2001). This is analogous
to a driver sitting in a cart pulled by a donkey and steering by dangling a carrot in the desired
direction in front of the donkey using a fishing-rod. The drawback of this, and most of the
other standard algorithms, is that vehicles have a tendency to take short cuts around curves,
which is inadvisable in a forest environment where trees and other obstacles are often situated
close to a defined route. As part of our research efforts, a new algorithm Follow the Past
(Hellström, Ringdahl 2005) has been developed. The idea is to utilize the driver’s steering
commands during a learning phase and compensate for any deviations from the route, which
12
may occur if the machine has avoided an obstacle or if the positioning system is not
sufficiently accurate. Follow the Past consists of three separate sub-functions:
1. Imitate the steering angles that the driver used during the learning phase.
2. Turn to travel in the same direction as during the learning phase
3. Turn towards the route if the machine is located beside it.
All of the sub-functions propose steering angles, which are then summed to obtain a value
that is used to steer the vehicle. The developed algorithm works well, and the vehicle follows
an intended route with good precision without taking short cuts around corners, provided that
there are no obstacles nearby. A module based on the algorithm VFH+ (Borenstein, Koren
1991) is responsible for avoiding obstacles and has a higher priority than the route-following
module. The steering angle is corrected if an obstacle is detected on the route. The vehicle
then waits for an intervention by an operator if too large a correction is needed to avoid
collision.
7. FIELD STUDY
The techniques for path-tracking, localization, and obstacle detection described above were
evaluated on a Valmet 830 forwarder equipped with a LMS221 laser scanner for obstacle
detection and a Javad Maxor RTK DGPS for estimating position and heading. To compensate
for the effects of the vehicle’s rolling and pitching on GPS positional data, an AHRS400 gyro
was used.
To assess the performance of the developed system in general and the path-tracking algorithm
Follow the Past in particular, a 160 metres long path was tracked four times while measuring
the distance to the reference path with the GPS. In addition, manual measurements were
obtained by marking the ground at the rear tire of the forest machine at intervals during each
run and then measuring the distance between the reference path and the new run with a
measuring tape. The tests were performed on flat ground to minimize the effect of vehicle roll
and pitch. The ground was covered by a thin layer of snow.
To test the autonomous vehicle’s ability to avoid obstacles, a loading pallet was placed
slightly to the left of the learnt path in one test, and slightly to the right of the path in another
test. To detect obstacles a laser scanner is used to generate pictures from which an occupancy
grid is derived, and the obstacle-avoidance algorithm VFH+ is used to modify the direction of
travel if any obstacles are detected along the planned path.
The average distance from the path in the four runs was 0.10 metres according to the GPS,
and 0.07 metres according to the manual measurements while the maximum errors were 0.35
metres and 0.4 metres, respectively (Table 1; Fig. 2). The differences in distances obtained by
the two measuring techniques are well within the margins of error, given the accuracy of the
GPS and the manual measuring procedure.
13
Table 1. Distances from the reference path presented as means (m) and standard deviations
(
σ
) for each run and for all measurements together. The values are presented separately for
the manual measurements (man) and the GPS measurements (gps). N is the number of
measurements in each run.
RUN
[
]
mm
man
[
]
m
man
σ
man
N
[
]
mm
gps
[
]
m
gps
σ
gps
N
1 0.09 0.09 30 0.10 0.09 548
2 0.09 0.08 27 0.10 0.09 590
3 0.06 0.05 29 0.10 0.09 593
4 0.07 0.06 26 0.10 0.09 647
average 0.07 0.07 - 0.10 0.09 -
The probability of path-tracking errors less than 23 centimetres is 90% (Figure 3). Errors
larger than 38 centimetres were never observed during any of the four test runs, in which a
total of 2378 measurements were obtained.
With the implemented obstacle avoidance algorithm VFH+, the vehicle is able to avoid
obstacles and subsequently return to the learnt path (Figure 4). The obstacles were detected
using a laser scanner and an occupancy grid.
-30 -20 -10 0 10 20
-35
-30
-25
-20
-15
-10
-5
0
5
[m]
[m]
Reference path
Run1
Run2
Run3
Run4
14
Figure 2. Comparison of a 160-metre long reference path and four autonomous runs with the
developed path-tracking algorithm Follow the Past. The data show that the runs coincide
almost completely.
Figure 3. Cumulative distribution function (CDF) for the path-tracking errors (distance from
the reference path), obtained using the GPS.
Figure 4. Obstacle avoidance. (Left) The obstacle is placed to the left of the centreline of the
vehicle. (Right) The obstacle is placed to the right of the centreline of the vehicle. The
vehicle’s starting point is at the upper left corner of the figure. The obstacles were detected
using a laser scanner and an occupancy grid.
8. DISCUSSION AND CONCLUSIONS
The presented system has demonstrated both possibilities and difficulties associated with
autonomous forest machines. It is shown that it is quite possible for them to learn and track a
path previously demonstrated by an operator. The overall performance was very good in all
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0
10
20
30
40
50
60
70
80
90
100
Distance to reference path [m]
Cumulative Probalbility [%]
15
four tests. For the future, the field study should be repeated in ordinary forest terrain with
obstacles and slippery slopes in different directions, and under a tree canopy.
A new path-tracking algorithm has been developed to reduce deviations by utilizing the
driver’s steering commands. It would also be possible to use information from a map to define
a desired path, but further research is required before this becomes a realistic alternative to the
operator demonstrating a safe path through the forest.
To determine the vehicle’s position and heading, a highly accurate satellite navigation system
is used. The results from the field study confirm the centimetre accuracy of the GPS claimed
by the manufacturer. Although the costs of such systems will no doubt decline in the future,
there is a need to develop less expensive techniques for determining position and heading.
Another reason for developing such techniques is that the accuracy of satellite navigation
systems can deteriorate considerably when the GPS-signals are obstructed by obstacles such
as large trees. For these reasons we have developed algorithms for laser odometry, neural
networks to improve wheel odometry, and techniques for combining information provided by
several position and heading sensors. The laser odometry system is highly sensitive to
inaccurate readings, and tests in forest environments have shown that it is difficult to make it
more accurate than wheel odometry. The neural network is usually able to increase the time
that wheel odometry can be used before the errors grow too large. The gyro gives the most
accurate heading information after the satellite navigation system.
Detecting obstacles is crucial for an autonomous vehicle. This is quite difficult in a forest
environment because the uneven terrain gives many “false positive” readings, i.e. detecting
obstacles at times when the ground simply becomes visible to the sensors, for example just
before a steep incline. Distinguishing between an obstacle and something the machine can
simply drive over, e.g. a large stone versus a shrub, presents further challenges. Detecting
“negative obstacles”, e.g. a ditch or steep slope, is also a problem to consider. To detect
obstacles as reliably as possible, the system supports use of several different sensors, although
to date we have mainly used the laser scanner, which is able to detect obstacles in front of the
vehicle. To keep track of obstacles and reduce the number of false positive readings, an
occupancy grid based on Bayesian updating is used.
To avoid detected obstacles, the VFH+ algorithm is used. We have found that this works well
for avoiding obstacles, but performs less well when there are narrow passages to negotiate
with obstacles close to both sides of the vehicle
. The behaviour of the vehicle, in terms of
variables such as how close to an obstacle it will go, can be altered by changing parameters in
the VFH+ algorithm. Major efforts to develop systems for avoiding obstacles in rough terrain
are being made by various groups around the world (cf. Iagnemma & Boehler 2006.).
A crucial factor that will strongly influence if and when autonomous forest machines are
introduced in forestry is interest from machine manufacturers. Major efforts are required to
proceed from academic research to products with acceptable performance for the end users.
We estimate that this will take at least 10-20 years depending, of course, on the scale of
resources invested in such efforts by both academic institutions and the forest industry.
16
ACKNOWLEDGEMENTS
The study was financed by Kempe foundation, VINNOVA and SLU. Thanks to Sees Editing
Ltd for revising the English.
REFERENCES
Andersson, S. 2004. Skogsteknik förr och nu. The Swedish Society of Forest History.
Årsskrift 2004: 102-116. (In Swedish).
Anon 2007. Statistical Yearbook of Forestry. Sveriges officiella statistik, National Board of
Forestry, Jönköping, Sweden. p. 108, 121. (In Swedish with English summary).
Barton, M. J. 2001. Controller Development and Implementation for Path Planning and
Following in an Autonomous Urban Vehicle, PhD thesis, University of Sydney, Australia.
Bergkvist, I., Nordén, B & Lundström 2006, H. Bergkvist, I.; Nordén, B. & Lundström, H.
Innovative unmanned harvester system, Skogforsk Results No 2. Uppsala, Sweden.
Borenstein, J., Koren, Y. 1991, The Vector Field Histogram - Fast Obstacle Avoidance for
Mobile Robots, IEEE Transactions on Robotics and Automation 7(3):278-288.
Brown, N.C. 1949. Logging. The principles and methods of harvesting timber in the United
States and Canada. John Wiley & Sons, Inc., New York. Chapman & Hall, Limited, London.
Drushka, K. & Konttinen, H. 1997. Tracks in the Forest. The Evolution of logging machinery.
Timberjack group, Helsinki, Finland. 254 pp.
Ekman, W. 1908. Skogsteknisk handbok. C.E. Fritzes Bokförlags Aktiebolag.
Centraltryckeriet. Stockholm, Sweden. (In Swedish).
Fryk, J., Larsson, M., Myhrman, D. & Nordansjö, I. 1991. Forest Operations in Sweden.
Forskningsstiftelsen Skogsarbeten. Kista, Sweden. ISBN: 91 7614 076 8.
Gellerstedt, S., Asplund, C. & Wästerlund, I. 1996. Robotics with application to forestry.
Uppsatser och Resultat nr. 285. Department of Operational Efficiency, SLU, Garpenberg,
Sweden. ISSN: 0282-2377.
Golob, T.B. 1981. Worker and his tools. Seminar on Occupational Health & Safety and
Applied Ergonomics on Highly Mechanised Logging operations. FAO/ECE/ILO, Ottawa,
Canada.
Hallonborg U. 2003. Unmanned forestry machines can be competitive, Skogforsk Results No
4, Uppsala, Sweden
Hellström, T., Johansson T. & Ringdahl O. 2006. Development of an Autonomous Forest
Machine for Path Tracking, volume 25 of Springer Tracts in Advanced Robotics, pages 603-
614. Springer, Field and Service Robotics: results of the 5th international conference, 2006.
17
Hellström, T. & Ringdahl, O. 2005. Follow the Past: a path-tracking algorithm for
autonomous vehicles, Int. J. Vehicle Autonomous Systems, 4 (2-4):216–224.
Iagnemma, K. & Buehler, M. 2006. Journal of Robotic Systems, 2006, Volume 23, Issue 9,
Special Issue on the DARPA Grand Challenge. John Wiley and Sons Ltd. ISSN: 0741-2223.
Kourtz, P. 1996. Autonomous forestry robots for brushing and thinning in young conifer
stands: early Canadian experiences. In: Gellerstedt, S., Asplund, C. & Wästerlund, I. 1996.
(Eds.) Robotics with application to forestry. Uppsatser och Resultat nr. 285. Department of
Operational Efficiency, SLU, Garpenberg, Sweden. ISSN: 0282-2377.
Kurabyashi, D. & Asaman, H. 2001. Navigation method by distributed map management for
autonomous bush removal robot. In: Halme, A., Chatila, R. & Prassler, E. (Eds.) FSR2001
Proceedings of the 3rd International Conference on Field and Service Robotics, June 11-13,
2001, Finland. ISBN: 952-5183-13-0.
Lidén, E. 1995. Inquiries into harvesting systems, quantities, and machine ownership in
Swedish industrial forestry – status 1985/86 and 1992/93 and predictions for 1997/98.
Uppsatser och resultat 273, Department of operational efficiency, SLU, Garpenberg, Sweden.
ISSN: 0282-2377.
Malmberg, C-E. 1988. När skogsbruket började mekanisera, In. Anon 1988. Skogsteknisk
forskning och utveckling i Sverige under 50 år. Redogörelse nr 6. Forskningsstiftelsen
Skogsarbeten, Kista, Sweden (In Swedish).
Nordansjö, I. 1988. Teknisk utveckling i skogsbruket 1938-1988, In. Anon 1988,
Skogsteknisk forskning och utveckling i Sverige under 50 år. Redogörelse nr 6.
Forskningsstiftelsen Skogsarbeten, Kista, Sweden. (In Swedish).
Ringdahl, O. 2007. Autonomous Ground Vehicles in Unstructured Forest Environment,
Licentiate thesis, Department of Computing Science, Umeå University, 2007. ISBN-978-91-
7264-373-4.
Silversides, C.R. 1997. Broadaxe to flying shear. The mechanization of forest harvesting east
of the Rockies. National museum of science and technology. Transformation series 6. Ottawa,
Canada. 174 pp
Staaf, A. 1988. Brunetten – en uppmärksammad slutprodukt från VSA, In. Anon 1988.
Skogsteknisk forskning och utveckling i Sverige under 50 år. Redogörelse nr 6.
Forskningsstiftelsen Skogsarbeten, Kista, Sweden. (In Swedish).
Sundberg, U. 1978. Teknik i skog. I: Skogshögskolan 150 år. Problem och idéer i svenskt
skogsbruk 1828-1978. Liber förlag. Stockholm Sweden. ISBN 91-7088-979-1. (In Swedish).
Vestlund, K. 2005. Aspects of Automation of Selective Cleaning. Acta Universitatis
Agriculturae Sueciae. Doctoral Thesis No. 2005:74. Faculty of Forest Sciences.
Vestlund, K. Nordfjell, T. & Eliasson, L. 2005. Comparison of Human and Computer-based
Selective Cleaning. Silva Fennica 39(4): 509-523.
Östberg, M. 1990. En smedjas förvandling - ÖSAs historia. Nyströms tryckeri. Bollnäs,
Sweden. (In Swedish).
