Online Speed Adaptation Using Supervised Learning for High-Speed, Off-Road Autonomous Driving.
ABSTRACT The mobile robotics community has traditionally addressed motion planning and navigation in terms of steering decisions. However, selecting the best speed is also important - beyond its relationship to stopping distance and lateral maneuverability. Consider a high-speed (35 mph) autonomous vehi- cle driving off-road through challenging desert ter- rain. The vehicle should drive slowly on terrain that poses substantial risk. However, it should not daw- dle on safe terrain. In this paper we address one aspect of risk - shock to the vehicle. We present an algorithm for trading-off shock and speed in real- time and without human intervention. The trade-off is optimized using supervised learning to match hu- man driving. The learning process is essential due to the discontinuous and spatially correlated nature of the control problem - classical techniques do not directly apply. We evaluate performance over hun- dreds of miles of autonomous driving, including performance during the 2005 DARPA Grand Chal- lenge. This approach was the deciding factor in our vehicle's speed for nearly 20% of the DARPA com- petition - more than any other constraint except the DARPA-imposed speed limits - and resulted in the fastest finishing time. In mobile robotics, motion planning and navigation have tra- ditionally focused on steering decisions. This paper presents speed decisions as another crucial part of planning - beyond the relationship of speed to obstacle avoidance concerns, such as stopping distance and lateral maneuverability. Consider a high-speed (35 mph) autonomous vehicle driving off-road through challenging desert terrain. We want the vehicle to drive slower on more dangerous terrain. However, we also want to minimize completion time. Thus, the robot must trade-off speed and risk in real-time. This is a natural pro- cess for human drivers, but it is not at all trivial to endow a robot with this ability. We address this trade-off for one component of risk: the shock the vehicle experiences. Minimizing shock is impor- tant for several reasons. First, shock increases the risk of damage to the vehicle, its mechanical actuators, and its elec- tronic components. Second, a key perceptive technology, laser range scanning, relies on accurate estimation of orien- tation. Shock causes the vehicle to shake violently, making accurate estimates difficult. Third, shocks substantially re- duce traction during oscillations. Finally, we demonstrate that shock is strongly correlated with speed and, independently, with subjectively difficult terrain. That is, minimizing shock implies slowing on challenging roads when necessary - a cru- cial behavior to mitigate risk to the vehicle. Our algorithm uses the linear relationship between shock and speed which we derive analytically. The algorithm has three states. First, the vehicle drives at the maximum allowed speed until a shock threshold is exceeded. Second, the vehicle slows immediately to bring itself within the shock threshold using the relationship between speed and shock. Finally, the
- [Show abstract] [Hide abstract]
ABSTRACT: In this paper, we propose a speed control system for autonomous mobile robots using long-range road estimation. In our method, the long-range road is estimated using the images from a camera and the result of near-range road estimation from a laser scanner. Near-range road surface conditions are estimated by using information of remission value as reflectivity of a laser. Our speed control determines velocity from the width and distance of the road, and this speed control makes the robot's performance more efficient (speedy). Our proposed method is adapted to various environments and it is effective to moving objects. Experimental results show that efficient movement can be performed by our proposed method.01/2011;
- [Show abstract] [Hide abstract]
ABSTRACT: This paper addresses the problem of closing the loop from perception to action selection for unmanned ground vehicles, with a focus on navigating slopes. A new non-parametric learning technique is presented to generate a mobility repre- sentation where maximum feasible speed is used as a criterion to classify the world. The inputs to the algorithm are terrain gradients derived from an elevation map and past observations of wheel slip. It is argued that such a representation can aid in path planning with improved selection of vehicle heading and operating velocity in off-road slopes. Results of mobility map generation and its benefits to path planning are shown.The International Journal of Robotics Research 01/2010; 29:997-1018. · 2.86 Impact Factor
- [Show abstract] [Hide abstract]
ABSTRACT: Rough terrains pose a major threat for the stability and safety of autonomous vehicles especially at high speeds. The mobile robotics community has traditionally treated rough terrain segments as hazards to be avoided and models the risk involved in traversing such segments as a characteristic of terrain surface and vehicle design. However, the risk due to such terrain patches depends greatly on the vehicle speed and can be easily reduced by regulating this speed. In this paper, we address shock experienced by the vehicle which is a major aspect of the risk. We present a local planning approach that incorporates the relation between shock, speed and terrain roughness into its cost function, thus resulting in plans that reduce shock. Terrain roughness estimates are made based on the Difference of Normals technique. We experimentally validate our approach in a real world setup.Proceedings of Conference on Advances In Robotics, Pune, India; 01/2013
Online Speed Adaptation using Supervised Learning
for High-Speed, Off-Road Autonomous Driving
David Stavens, Gabriel Hoffmann, and Sebastian Thrun
Stanford Artificial Intelligence Laboratory
Stanford, CA 94305-9010
The mobile robotics community has traditionally
addressed motion planning and navigation in terms
of steering decisions. However, selecting the best
speed is also important – beyond its relationship
to stopping distance and lateral maneuverability.
Consider a high-speed (35 mph) autonomous vehi-
cle driving off-road through challenging desert ter-
rain. The vehicle should drive slowly on terrain that
poses substantial risk. However, it should not daw-
dle on safe terrain. In this paper we address one
aspect of risk – shock to the vehicle. We present an
algorithm for trading-off shock and speed in real-
time and without human intervention. The trade-off
is optimized using supervised learning to match hu-
man driving. The learning process is essential due
to the discontinuous and spatially correlated nature
of the control problem – classical techniques do not
directly apply. We evaluate performance over hun-
dreds of miles of autonomous driving, including
performance during the 2005 DARPA Grand Chal-
lenge. This approach was the deciding factor in our
vehicle’s speed for nearly 20% of the DARPA com-
petition – more than any other constraint except the
DARPA-imposed speed limits – and resulted in the
fastest finishing time.
In mobile robotics, motion planning and navigation have tra-
ditionally focused on steering decisions. This paper presents
speed decisions as another crucial part of planning – beyond
the relationship of speed to obstacle avoidance concerns, such
as stopping distance and lateral maneuverability. Consider
a high-speed (35 mph) autonomous vehicle driving off-road
through challenging desert terrain. We want the vehicle to
drive slower on more dangerous terrain. However, we also
want to minimize completion time. Thus, the robot must
trade-off speed and risk in real-time. This is a natural pro-
cess for human drivers, but it is not at all trivial to endow a
robot with this ability.
We address this trade-off for one component of risk: the
shock the vehicle experiences. Minimizing shock is impor-
tant for several reasons. First, shock increases the risk of
damage to the vehicle, its mechanical actuators, and its elec-
tronic components. Second, a key perceptive technology,
laser range scanning, relies on accurate estimation of orien-
tation. Shock causes the vehicle to shake violently, making
accurate estimates difficult. Third, shocks substantially re-
ducetractionduringoscillations. Finally, wedemonstratethat
shock is strongly correlated with speed and, independently,
with subjectively difficult terrain. That is, minimizing shock
implies slowing on challenging roads when necessary – a cru-
cial behavior to mitigate risk to the vehicle.
Our algorithm uses the linear relationship between shock
and speed which we derive analytically. The algorithm has
three states. First, the vehicle drives at the maximum allowed
speeduntilashockthresholdisexceeded. Second, thevehicle
slows immediately to bring itself within the shock threshold
using the relationship between speed and shock. Finally, the
vehicle gradually accelerates. It returns to the first state, or
the second if the threshold is exceeded during acceleration.
To maximize safety, the vehicle must react to shock
by slowing down immediately, resulting in a discontinuous
speed command.Further, it should accelerate cautiously,
because rough terrain tends to be clustered, a property we
demonstrate experimentally. For these reasons, it is not pos-
sible to determine the optimal parameters using classical con-
trol techniques. The selection of parameters cannot be tuned
by any physical model of the system, as the performance is
best measured statistically by driving experience.
Therefore, we generate the actual parameters for this al-
gorithm – the shock threshold and acceleration rate – using
supervised learning. The algorithm generates a reasonable
match to the human teacher. Of course, humans are proac-
tive, also using perception to make speed decisions. We use
inertial-only cues because it is very difficult to assess terrain
roughness with perceptive sensors. The accuracy required ex-
ceeds that for obstacle avoidance by a substantial margin be-
cause rough patches are often just a few centimeters in height.
Still, experimental results indicate our reactive approach is
very effective in practice.
Our algorithm is part of the software suite developed for
Stanley, an autonomous vehicle entered by Stanford Univer-
sity in the 2005 DARPA Grand Challenge [DARPA, 2004].
Stanley is shown in Figure 1. The Grand Challenge was a
robot race organized by the U.S. Government. In 2005, ve-
hicles had to navigate 132 miles of unrehearsed desert ter-
Figure 1: Stanley won the 2005 DARPA Grand Challenge by
completing a 132 mile desert route in just under 7 hours. The
algorithm described in this paper was the deciding factor in
Stanley’s speed for nearly 20% of the race – more than any
other constraint except the DARPA-imposed speed limits.
rain. The algorithm described here played an important role
in that event. Specifically, it was the deciding factor in our
robot’s speed for nearly 20% of the competition – more than
any other constraint except the DARPA imposed speed limit.
It resulted in the fastest finishing time of any robot.
There has been extensive work on terramechanics [Bekker,
1956; 1969; Wong, 1989], the guidance of autonomous vehi-
cles through rough terrain. That work generally falls into two
categories. One focuses on high-speed navigation. However,
it requires that terrain maps and topology be known before-
hand [Shimoda et al., 2005; Spenko et al., 2006]. The other
analyzes terrain ruggedness in an online fashion by driving
over it[Sadhukhan et al., 2004; Brooks and Iagnemma, 2005;
Iagnemma et al., 2004]. However, vehicle speed is very slow,
less than 1.8 mph. Further, classification focused on ter-
rain type, not roughness. Our work combines the strengths
of both categories: online terrain roughness assessment and
Successful obstacle detection systems have been built for
high-speed autonomous driving [Thrun et al., 2006b; 2006a;
Urmson et al., 2006; 2004; Kelly and Stentz, 1998]. How-
ever, as we mentioned in Section 1, the precision needed for
determining terrain ruggedness exceeds that required for ob-
stacle avoidance by a substantial margin. Thus, while these
systems are excellent at detecting and avoiding static obsta-
cles, they do not protect the vehicle from the effects of road
Some work has explicitly considered speed as important
for path planning. For example, [Fox et al., 1996]presents a
method for trading off “progress” and “velocity” with regard
to the number and type of obstacles present. That is, slower
speeds are desirable in more cluttered environments. Faster
speeds are better for open spaces. This concept, applied to
desert driving, is analogous to slowing for turns in the road
and for discrete obstacles. Our vehicle exhibits those behav-
iors as well. However, they are separate from the algorithm
we describe in this paper.
Other entrants in the 2005 DARPA Grand Challenge re-
alized the importance of adapting speed to rough terrain.
Figure 2: We find empirically that the relationship between
perceived vertical acceleration amplitude and vehicle speed
in the operational region of interest.
For example, two teams from Carnegie Mellon University
(CMU) used “preplanning” for speed selection [Gutierrez et
al., 2005; Urmson et al., 2006; 2004; NOVA, 2006]. For
months, members of the CMU teams collected extensive en-
vironment data in the general area where the competition was
thought to be held. Once the precise route was revealed by
DARPA, two hours before the race, the team preprogrammed
speeds according to the previously collected data. The end-
result of the CMU approach is similar to ours. However, our
fully-online method requires neither human intervention nor
prior knowledge of the terrain. The latter distinction is par-
ticularly important since desert terrain readily changes over
time due to many factors, including weather.
Our algorithm is derived in three steps. First, we derive the
linear relationship between speed and shock. Second, we
leverage this relationship to build an algorithm for determin-
ing an appropriate speed. Finally, we tune the parameters of
the algorithm using supervised learning.
Speed Selection Algorithm
The mapping between vehicle speed and shock experienced
is fundamental to our technique. Here, shock is measured by
the amplitude of acceleration in the vertical, z-direction, of
the vehicle’s main body. The vehicle’s speed is v and posi-
tion along the path is p. The ground varies in height, zg(p),
pushing stiffly on the car’s tires. The ground is modeled to
be the sum of sinusoids of a large number of unknown spatial
frequencies and amplitudes. Consider one component of this
summation, with a spatial frequency of ωs, and an amplitude
Az,g,ωs. Due to motion of the vehicle, the frequency in time
of zg,ωsis ω = vωs. By taking two derivatives, the accelera-
tion, ¨ zg,ωs, has an amplitude of v2ω2
of vωs. This is the acceleration input to the vehicle’s tire.
Z Acceleration Mapping and Filtering
sAz,g,ωsand a frequency
Figure 3: Filtered vs. unfiltered IMU data. The 40-tap FIR
band pass filter extracts the excitation of the suspension sys-
tem, in the 0.3 to 12 Hz band. This removes the offset of
gravity, as terrain slope varies, and higher frequency noise,
such as engine system vibration, that is unrelated to oscilla-
tions in the vehicle’s suspension.
To model the response of the suspension system, we use
the quarter car model [Gillespie, 1992]. The tire is modeled
as a connection between the axle and the road with a spring
with stiffness kt, and the suspension system is modeled as a
connection between the axle and the vehicle’s main body, us-
Using this model, the amplitude of acceleration of the main
body of the vehicle can be easily approximated, by assuming
infinitely stiff tires, to be
(csvωs)2+ (ks− mv2ω2
where m is the mass of one quarter of the vehicle, and the
magnitude of the acceleration input to the tire is taken to be
a standard automobile is around 1 to 1.5 Hz. Below this fre-
quency, the amplitude increases from zero at 0 Hz, slowly, but
exponentially, with frequency. Above this frequency, the am-
active damping systems further quench the effect of the reso-
nant frequency, leading to a curve that is nearly linear in fre-
quency throughout, to within the noise of the sensors, which
also measure the effect of random forcing disturbances.
The frequency, ω, is directly proportional to speed v, so
the amplitude of acceleration response of the vehicle is also
approximately a linear function of the velocity. Now, by su-
perposition, considering the sum of the responses for all com-
ponents of the terrain profile at all spatial frequencies, the am-
plitude of the acceleration response of the vehicle, A¨ z, is the
sum of many approximately linear functions in v, and hence
is approximately linear in v itself. This theoretical analysis
can also be verified experimentally as in Figure 2.
This approximation assumed infinitely stiff tires. In reality,
the finite stiffness of tires leads to a higher frequency reso-
A¨ z,ωs= v2ω2
sAz,g,ωs, as described above. The resonant frequency of
nance in the suspension system around 10 to 20 Hz, followed
by an exponential drop off in A¨ z, to zero.
Although the acceleration is approximately a linear func-
tion of v, a barrier to the use of pure z-accelerometer inertial
measurement unit (IMU) data is that acceleration can reflect
other inputs than road roughness. First, the projection of the
gravity vector onto the z-axis, which changes with the pitch
of the road, causes a large, low frequency offset. Second,
the driveline and non-uniformities in the wheels cause higher
To remove these effects, we pass the IMU data through a
40-tap, FIR band pass filter to extract the excitation of the
suspension system, in the 0.3 to 12 Hz band. This band was
found, experimentally, to both reject the offset of gravity and
eliminate vibrations unrelated to the suspension system, with-
out eliminating the vibration response of the suspension.
A comparison of the raw and filtered data is shown in Fig-
ure 3. The mean of the raw data, acquired on terrain with no
slope, is about 1G, whereas the mean of the filtered data is
about 0G. Therefore, solely for presentation in this figure, 1G
was subtracted from the raw data.
Our speed selection algorithm (the velocity controller) has
three states. Each road is assumed to have a speed limit (γ, in
mph) that represents the upper bound on vehicle speed along
that road. First, the vehicle travels at the speed limit until
a road event generates a shock that exceeds an acceptable
threshold (α, in Gs). Second, it reduces speed such that the
event would have been less than α. Finally, it accelerates (at
rate β, in mph/s) back to the speed limit.
More formally, recall that, for a given rough segment of
road, the shock experienced by the vehicle is linear with re-
spect to speed. Every dt seconds, the vehicle takes a reading
at its current location p along the route. Filtered shock ¨ zpand
speed vpare observed. The optimal speed at p should have
Generating Speed Recommendations
That is, v∗
imum acceptable shock (α) to the vehicle.
Our instant recommendation for vehicle speed is v∗
tice that the vehicle is slowing reactively. Reactive behavior
has no effect if shock is instant, random, and not clustered.
That is, our approach relies on one shock event “announcing”
the arrival of another, momentarily. Experimental results on
hundreds of miles of desert indicate this is an effective strat-
We consider v∗
rates these observations over time, generating the actual rec-
ommended speed, vr
pis the velocity that would have delivered the max-
pan observation. The algorithm incorpo-
pseries a velocity plan.
For our purposes, we also clamp vr
than 5 mph. We call the vr
psuch that it is never less
Figure 4: The results of coordinate descent machine learning
are shown. The learning generated the race parameters of
.25Gsand1mph/s, agoodmatchtohumandriving. Notethat
the human often drives under the speed limit. This suggests
the importance of our algorithm for desert driving.
We use supervised machine learning to select α and β. Some
might argue we should select these parameters using explicit
hardware lifetime calculations rather than matching a human
teacher. However, an autonomous vehicle has hundreds of
components. Analyzing the effect of shock on each would
require a great deal of time. It would also be costly – each
component would have to break before we could determine
its limit. We believe that matching a human’s speed provides
a useful and more efficient bound on acceptable shock.
The velocity plan generated by a specific parameterization
is scored against a human driver’s velocity decisions at all p.
The parameters are then iteratively modified according to a
scoring function and coordinate descent.
objective/scoring function we wish to minimize is
Supervised Machine Learning
pbe the speed of the human driver at position p. The
???1 + αβ−1?
where ψ = 1 if vr
penalizes speeding relative to the human driver. The value 3
was selected arbitrarily. The αβ−1term penalizes parameter-
izations that tolerate a great deal of shock or accelerate too
For this experiment, a human driver drove 1.91 miles of
the previous year’s (2004) Grand Challenge route. The vehi-
cle’s state vector was estimated by an Unscented Kalman Fil-
ter (UKF) [Julier and Uhlmann, 1997]. The UKF combined
measurements from a Novatel GPS with Omnistar differen-
tial corrections, a GPS compass, the IMU, and wheel speed
encoders. The GPS devices produced data at 10Hz. All other
devices and the UKF itself produced data at 100Hz. All data
is logged with millisecond accurate timestamps.
pand ψ = 3 otherwise. The ψ term
We do not model driver reaction time, and we correct for
the delay due to IMU filtering. Thus, in this data, driver ve-
locity at position p is the vehicle’s recorded velocity at po-
sition p, according to the UKF. The algorithm’s velocity at
position p is the decision of the velocity controller consider-
ing all data up to and including the IMU and UKF data points
logged at position p.
The parameterization learned was (α, β) = (.27 Gs,
.909 mph/s). This is illustrated in Figure 4. The top and
bottom halves of the figure show the training and test sets,
respectively. The horizontal axis is position along the 2004
Grand Challenge route in miles. The vertical axis is speed
in mph. For practical reasons we rounded these parameters
to .25 and 1, respectively, which are essentially identical in
terms of vehicle performance. The robot used these parame-
ters during the Grand Challenge event and they are evaluated
extensively in future sections.
We notice a reasonable match to human driving. Without
the penalty for speeding, a closer match is possible. However,
that parameterization causes the vehicle to drive too aggres-
sively. This is, no doubt, because a human driver also uses
perceptual cues to determine speed.
Notice that the speed limits are often much greater than
the safe speed the human (and the algorithm) selects. This
demonstrates the importance of our technique over blindly
following the provided speed limits. Notice also, however,
both the human and the algorithm at times exceed the speed
limit. This is not allowed on the actual robot or in the experi-
mentsinSection4. However, wepermititheretosimplifythe
scoring function, the learning process, and the idea of match-
ing a human driver’s speed.
In our first experiment, we evaluate the algorithm on sections
of off-road driving that we identify (qualitatively) as partic-
ularly easy and, alternatively, as challenging. The purpose
of this experiment is to spot-check performance. We seek to
verify that our algorithm, as tuned by the learning, does not
drive too slowly on easy terrain or too fast on difficult terrain.
We select two seven mile stretches from the 2005 Grand
Challenge route. The first stretch includes miles 0-7. This is
very easy terrain consisting of wide dirt roads and dry lakes.
The second includes miles 122-129. This terrain is very chal-
lenging: a narrow mountain road with a 200 ft drop on one
side. Both sections were graded (smoothed with a large plow)
not long before the Grand Challenge event. Based on video
collected by the vehicle and by DARPA, the mountain road
is clearly much more challenging to drive. However, neither
has obviously greater potential for vertical shock.
Figure 5 presents results from the first seven miles of the
course – the easy example. In the figure, points represent a
single shock event along the route. The vertical axis plots
shock the vehicle experiences while driving the route with
the velocity controller. The horizontal axis plots shock expe-
rienced driving the speed limits alone. We notice two things.
First, shock above our .25G threshold is very rare. This re-
flects on the particularly straightforward driving conditions.
Figure 5: Each point indicates a specific shock event with
both the velocity controller (vertical axis) versus speed limits
alone (horizontal axis). There are very few high shock events
on this easy terrain. The velocity controller does not reduce
shock or speed much.
Second, we notice that the shock experienced with the speed
limits alone is, for all events, roughly the same as that expe-
rienced with the velocity controller. Thus, because shock is
linear with speed, the velocity controller did not significantly
reduce speed on this easy terrain. This supports the idea that
the algorithm, with the learned parameters, does not cause
the vehicle to drive too slowly on easy terrain.
the same type of plot, with the same axis ranges, as Figure 5.
However, here, we plot data from the more difficult segment,
miles 122-129. In this terrain, poor perceptive sensing or loss
of traction – both due to excessive shock – could result in to-
tal vehicle loss. The velocity controller is very active, lower-
ing vehicle speed. The lower speed results in reduced shock,
better traction, and generally more favorable conditions for
successful navigation of this challenging terrain.
Figure 6 is
Having verified the basic function of the algorithm, we now
proceed to analyze its performance over long stretches of
driving. For long stretches, it becomes intractable to com-
pare algorithm performance to human decisions (as in Sec-
tion 3.3) or relative to qualitatively assessed terrain condi-
tions (as in Section 4.1). Therefore, we must consider other
metrics. Over long stretches of driving, we use the aggregate
reduction of large shocks to measure algorithm effectiveness.
However, to evaluate the reduction of large shocks, we can-
not simply sum the amount of shock. Extreme shock is rare.
Less than .3% of all readings (2005 Grand Challenge, Speed
Limits Alone) are above our maximum acceptable shock (α)
threshold. Thus, the aggregate effect of frequent, small shock
could mask the algorithm’s performance on rare, large shock.
To avoid this problem, we take the 4th power (“L4 Metric”)
of a shock before adding it to the sum. This accentuates large
shocks and diminishes small ones in the aggregate score.
Figure 6: Each point indicates a specific shock event with
both the velocity controller (vertical axis) versus speed limits
alone (horizontal axis). There are many high shock events
on this difficult terrain. The velocity controller reduces shock
and speed a great deal.
Whenever the vehicle travels along a road – regardless of
whether under human or autonomous control – it can map the
road for velocity control experiments. Each raw IMU read-
ing is logged along with the vehicle’s velocity and position
according to the UKF. Then, offline, each IMU reading can
be filtered and divided by the vehicle’s speed. This generates
a sequence of position tagged road roughness values in the
units Gs-felt-per-mile-per-hour. We use these sequences in
this section to evaluate the velocity controller.
We calculate completion time by summing the time spent
traversing each route segment (ie: from pito pi+1) accord-
ing to the average speed driven during that segment. 1/dt
is 100Hz. Because the velocity controller may change speed
instantly at each pi, we track the velocities it outputs – for
these simulations only – with a simple controller. It restricts
the maximum amount of change in velocity between succes-
sive pi’s, approximating vehicle capabilities. The limits are
.02 mph for increases and .09 mph for decreases. We use the
controller output for completion time and shock calculations.
Time vs. Shock for Four Desert Routes
We consider the trade-off of shock versus completion time on
four desert routes. See Figure 8. The horizontal axis plots
completion time. The vertical axis plots shock felt. Both are
for the algorithm normalized against speed limits alone. Each
of the four curves represents a different desert route: the 2005
Grand Challenge route, the 2004 Grand Challenge route, and
two routes outside of Phoenix, Arizona, USA that we created
for testing purposes. Individual points on the curves depict
the completion time versus shock trade-off for a particular set
of (α, β) values. For these curves, α is fixed and β is varied.
The asterisk on each curve indicates the trade-off generated
by the parameters selected in Section 3.3. The fifth aster-
isk, at the top-left, indicates what happens with speed limits
alone. We notice that, for all routes, completion time is in-
creased between 2.5% and 5% over the baseline. However,
Figure 7: Here we show where the algorithm reduced speed during the 2005 Grand Challenge event. The velocity controller
modifiedspeedduring 17.6% ofthe race. Height of theellipses shownatthe top isproportional to themaximum speed reduction
for each 1/10th mile segment. Numbers represent mile markers. Simulated terrain at the bottom approximates route elevation.
Figure 8: Here we plot the trade-off of completion time and
shock on four desert routes. α is constant (.25) and β is varied
along each curve. With the single set of parameters learned
in Section 3.3, the algorithm reduces shock by 50% to 80%
while increasing completion time by 2.5% to 5%.
shock is reduced by 50% to 80% compared to the baseline, a
It appears that the algorithm is most effective along the
2004 Grand Challenge route.
this. First, we prepared for the 2005 event along the 2004
course. Therefore, our algorithm is especially optimized for
that route. However, the effect is mostly due to a different
The 2004 course had sections that were in extremely bad
condition when this data was collected (June 2005). Such ex-
treme off-road driving was not seen on the other three routes.
Because the core functionality of our algorithm involves be-
ing cautious about clustered, dangerous terrain, it was par-
ticularly effective for the 2004 course. This implies our ap-
proach may be more beneficial as terrain difficulty increases.
Finally, we notice occasional abrupt, unsmooth drops in
shock in the figure. This is the result of a parameterization
that is just aggressive enough to slow the vehicle before a
major rough section of road. Using the parameterization im-
mediately prior did not slow the vehicle in time.
There are two reasons for
In Figure 7, we indicate where the algorithm reduced speed
during the 2005 Grand Challenge. Numbers and vertical lines
are mile markers. We depict terrain at the bottom that approx-
imates elevation data from the route. At the top of the figure,
ellipse height indicates the maximum speed reduction gener-
ated by the algorithm for that portion of the route. For the
purposes of this figure, the route was divided into 1/10th mile
segments. Log files taken by the vehicle during the Grand
Challenge event indicate the velocity controller was control-
ling vehicle speed for 17.6% of the entire race – more than
any other factor except speed limits.
We notice that the algorithm was particularly active in the
mountains and especially inactive in plains. We also notice
that miles 122-129 – which we identified in Section 4.1 as
the most challenging part of the route – do not represent the
greatest speed reductions the velocity controller made. This
is because speed limits (given by DARPA) were already low
in those regions, and the amount of speed reduction is, of
course, related to the accuracy of the original speed limit as a
good speed guide.
Results from the 2005 Grand Challenge
Speed decisions are a crucial part of planning – even beyond
their relationship to obstacle avoidance and lateral maneuver-
ability. We have presented an online approach to speed adap-
tation for high-speed off-road autonomous driving. Our ap-
proach addresses shock, an important component of overall
risk when driving off-road. Our method requires no rehearsal
or remote imaging of the route.
The algorithm has three states. First, the vehicle drives
at the speed limit until an acceptable shock threshold is ex-
ceeded. Second, the vehicle reduces speed to bring itself
within the threshold. Finally, the vehicle gradually increases
speed back to the speed limit. The algorithm uses the lin-
ear relationship between speed and shock to determine the
amount of speed reduction needed to stay within the limit.
This limit and the acceleration parameter from the third state
are tuned using supervised learning to match human driving.
Experimental results allow us to draw numerous conclu-
sions. First, our algorithm seems to be more cautious on
more difficult terrain, although this is difficult to prove in
the general case (Figures 5 and 6). Second, by slowing, we
can substantially reduce high shock events with minimal ef-
fect on route completion time (Figure 8). Finally, the algo-
rithm had significant influence on our vehicle’s behavior dur-
ing the 2005 Grand Challenge (Figures 5, 6, 7, and 8). It
reduced shock by 52% and slowed the vehicle in difficult ter-
rain, all while adding less than 10 minutes to the comple-
tion time. The algorithm also had more influence than any
other factor – except speed limits – on our robot’s speed.
Finally, it generated the fastest completion time despite be-
ing fully automated, unlike some other competitors in the
Grand Challenge[Gutierrez et al., 2005; Urmson et al., 2006;
David Stavens’ doctoral study is supported by the David
Cheriton Stanford Graduate Fellowship (SGF). The au-
thors gratefully acknowledge the Stanford Racing Team who
helped make this work possible, especially Michael Monte-
merlo, Hendrik Dahlkamp, and Sven Strohband.
[Bekker, 1956] G. Bekker. Theory of Land Locomotion. Uni-
versity of Michigan, 1956.
[Bekker, 1969] G. Bekker. Introduction to Terrain-Vehicle
Systems. University of Michigan, 1969.
[Brooks and Iagnemma, 2005] C.A. Brooks and K. Iag-
nemma. Vibration-based terrain classification for plane-
tary exploration rovers. IEEE Transactions on Robotics,
[DARPA, 2004] DARPA. DARPA Grand Challenge Rule-
book, 2004. http://www.darpa.mil/grandchallenge05/.
[Fox et al., 1996] D. Fox, W. Burgard, and S. Thrun. Con-
trolling synchro-drive robots with the dynamic window
approach to collision avoidance. In Proceedings of the
IEEE/RSJ International Conference on Intelligent Robots
and Systems, 1996.
[Gillespie, 1992] T. Gillespie. Fundamentals of Vehicle Dy-
namics. Society of Automotive Engineers, 1992.
[Gutierrez et al., 2005] A.Gutierrez, T. Galatali, J. P.Gonza-
lez, C. Urmson, and W. L. Whittaker. Preplanning for high
performance autonomous traverse of desert terrain exploit-
ing a priori knowledge to optimize speeds and to detail
paths. Technical Report CMU-RI-TR-05-54, Robotics In-
stitute, Carnegie Mellon University, Pittsburgh, PA, De-
[Iagnemma et al., 2004] K. Iagnemma, S. Kang, H. Shibly,
and S. Dubowsky. On-line terrain parameter estimation
for wheeled mobile robots with application to planetary
rovers. IEEE Transactions on Robotics and Automation,
2004. To appear.
[Julier and Uhlmann, 1997] S. Julier and J. Uhlmann. A new
extension of the Kalman filter to nonlinear systems. In
International Symposium on Aerospace/Defense Sensing,
Simulate and Controls, Orlando, FL, 1997.
[Kelly and Stentz, 1998] A. Kelly and A. Stentz. Rough ter-
rain autonomous mobility, part 2: An active vision, pre-
dictive control approach. Autonomous Robots, 5:163–198,
[NOVA, 2006] NOVA.
The Great Robot Race (Doc-
umentary of the DARPA Grand Challenge),
[Sadhukhan et al., 2004] D. Sadhukhan, C. Moore, and
E. Collins. Terrain estimation using internal sensors. In
Proceedings of the 10th IASTED International Conference
on Robotics and Applications (RA), Honolulu, Hawaii,
[Shimoda et al., 2005] S. Shimoda, Y. Kuroda, and K. Iag-
nemma. Potential fieldnavigationof highspeedunmanned
ground vehicles in uneven terrain. In Proceedings of the
IEEE International Conference on Robotics and Automa-
tion (ICRA), Barcelona, Spain, 2005.
[Spenko et al., 2006] M. Spenko, Y. Kuroda, S. Dubowsky,
and K. Iagnemma. Hazard avoidance for high-speed mo-
bile robots in rough terrain. Journal of Field Robotics,
[Thrun et al., 2006a] S.
A. Aron.Probabilistic terrain analysis for high-speed
desert driving. In G. Sukhatme, S. Schaal, W. Burgard,
and D. Fox, editors, Proceedings of the Robotics Science
and Systems Conference, Philadelphia, PA, 2006.
[Thrun et al., 2006b] S. Thrun,
H. Dahlkamp, D. Stavens, A. Aron, J. Diebel, P. Fong,
J. Gale, M. Halpenny, G. Hoffmann, K. Lau, C. Oakley,
M. Palatucci, V. Pratt, P. Stang, S. Strohband, C. Dupont,
L.-E. Jendrossek, C. Koelen, C. Markey, C. Rummel,
J. van Niekerk, E. Jensen, P. Alessandrini, G. Bradski,
B. Davies, S. Ettinger, A. Kaehler, A. Nefian, and P. Ma-
honey. Stanley: The robot that won the darpa grand
challenge. Journal of Field Robotics, 2006.
[Urmson et al., 2004] C. Urmson, J. Anhalt, M. Clark,
T. Galatali, J.P. Gonzalez, J. Gowdy, A. Gutierrez, S. Har-
baugh, M. Johnson-Roberson, H. Kato, P.L. Koon, K. Pe-
terson, B.K.Smith, S.Spiker, E.Tryzelaar, andW.L.Whit-
taker. High speed navigation of unrehearsed terrain: Red
team technology for grand challenge 2004.
Report CMU-RI-TR-04-37, Robotics Institute, Carnegie
Mellon University, Pittsburgh, PA, 2004.
[Urmson et al., 2006] C. Urmson, J. Anhalt, D. Bartz,
M. Clark, T. Galatali, A. Gutierrez, S. Harbaugh, J. John-
ston, H. Kato, P. Koon, W. Messner, N. Miller, A. Mosher,
K. Peterson, C. Ragusa, D. Ray, B. Smith, J. Snider,
S. Spiker, J. Struble, J. Ziglar, and W. L. Whittaker. A
robust approach to high-speed navigation for unrehearsed
desert terrain. Journal of Field Robotics, 23(8):467–508,
[Wong, 1989] J. Wong. Terramechanics and Off-Road Vehi-
cles. Elsevier, 1989.