- Email: [email protected]
Intelligent Sensors and Control for Commercial Vehicle Automation
Copyright © IFAC Advances in Automotive Control, Mohican State Park, Loudonville, Ohio, USA, 1998
INTELLIGENT SENSORS AND CONTROL FOR COMMERCIAL VEmCLE AUTOMATION
loannis Kanellakopoulos
PbyUis Nelson Osc:ar Stafsudd
UCLA Electrical Engineering, Los Angeles, CA 90095-1594 Abstract: One of the potentially most prominent applications of Intelligent Transportation Systems technology is the partial or full automation of commercial vehicles such as trucks and buses for freight and passenger transport. The design issues which are particular to these vehicles have largely been ignored thus far, especially issues related to sensor and actuator limitations. Here we address the longitudinal control problem for commercial heavy vehicles, and we present a new class of ranging sensors and a new class of fueVbrake controllers which eliminate or bypass many of these limitations, and can be used in all stages of ITS deployment. Copyright© 1998 IFAC Keywords: Commercial vehicles, NHS, Ranging sensors, Intelligent systems, Longitudinal control, Nonlinear control
1. INfRODUCTION
manding tasks than autonomous systems, such as coordinated driving in a group, but their time to commercialization is likely to be longer. 3. Automated highway systems add information obtained from the roadway infrastructure, such as messages regarding traffic conditions and road geometry, and lateral information from magnetic nails or reflective guardrails installed on the highway. Such systems can perform even more demanding tasks, like fully automated driving in a platoon, but must face many more obstacles (standardization, liability issues, public acceptance) on their way to implementation.
Intelligent Transportation Systems (ITS) technology is progressing at an ever-increasing rate, with exciting developments in all fronts, from driver information and assistance systems to Automated Highway Systems (AHS). This presentation discusses potential scenarios with varying levels of automation for commercial vehicles, and introduces a new sensor technology called IRIS (Intelligent Ranging with Infrared Sensors) and new nonlinear control algorithms which can be combined with existing sensors and actuators to produce economically feasible automation solutions.
Commercial vehicles, in particular, will reap significant benefits from all stages of automation. Collision warning systems increase safety and reduce accidents, while adaptive cruise control enhances driver comfort and reduces fuel consumption and emissions. Fleet operators can further reduce their costs using cooperative scenarios like the electronic towbar, in which one manually driven vehicle is followed by two or three driverless automated vehicles. Finally, in automated highway systems, fully automated vehicles will be able to carry freight and passengers with significantly enhanced safety, increased fuel efficiency, and much more predictable travel times.
One of the main points of contention in the ITS community is the level of vehicle-to-vehicle and vehicleto-roadway cooperation that can be assumed in current research. In that respect, systems currently in various stages of research and development can be classified into three categories: 1. Autonomous systems depend only on information obtained by the sensors located on the vehicle itself, usually relative distance and velocity to stationary objects and moving vehicles. They are therefore implementable in the immediate future, and in fact have started to appear as commercial products (collision warning, adaptive cruise control).
Two critical automation components in any of these scenarios, especially for commercial vehicles, are the sensors and the actuators. Existing sensor technologies can only provide the reliability required for these
2. Cooperative systems add information transmitted by neighboring vehicles, usually acceleration and steering inputs. Hence, they can perform more de-
79
of the preceding vehicle and the surrounding objects scatter the incident laser beam and reflect only a negligible fraction back to the receiver to be recorded on the second (laser on) image. There are, however, some surfaces which reflect most of the incident laser beam back to its source: these are the passive reflective patches molded into the taillights of all modern vehicles, as well the reflective paint used in the license plates of many states, which act as efficient retroreflectors in the illuminator's wavelength. Subtraction of the first (laser off) image from the second (laser on) one yields a clear sharp image of these reflective surfaces against a flat background. The distance to the preceding vehicle is then easily computed from the resulting clear pattern via standard triangulation schemes.
applications by increasing the cost to levels beyond those acceptable to the potential customers. Actuators, on the other hand, have built-in saturations and delays which significantly complicate the control task. These issues are addressed by our current research, in which we are developing (1) very inexpensive and highly accurate ranging sensors that can be used as stand-alone sensors for adaptive cruise control and vehicle following, and also in combination with radar or vision for improved collision warning, lane tracking, as well as driving in platoons, and (2) novel nonlinear control algorithms which can deal with the severe actuator saturations and delays present in commercial vehicles to yield good performance in both autonomous and cooperative scenarios. The reason for focusing on these projects is that we are interested in results that will be useful in all stages of automation and will potentially impact product development ranging from today's collision warning and adaptive cruise control to the future's fully automated AHS vehicles.
It is particularly important to note that even in this simple example the clutter is essentially nonexistent. This is due to the combined effect of (1) the narrow bandpass filter, which significantly reduces the contribution of broadband sources such as the sun or artificial lights without affecting the laser returns, and (2) the image subtraction process, which eliminates all returns except those produced by laser illumination.
2. IRIS: INTELLIGENT RANGING USING INFRARED SENSORS
With the modulated illuminator and synchronous detection of the IRIS system, the retroreflectors and license plate are the dominant features in the subtracted image. The IRIS computational requirements are thus significantly lower than those of a conventional approach, which would use two-dimensional image processing techniques to extract the position of the vehicle from the cluttered scene of the original image.
2.1 Operating principle Reliable ranging between adjacent vehicles is one of the key enabling technologies for ITS. However, the dynamically changing random environment of highway traffic makes it very difficult for existing technologies to achieve the necessary accuracy and reliability. In the case of radar and sonar, this is due to their susceptibility to secondary reflections, while vision-based systems require massive computations. Furthermore, most of these technologies are based on component parts which are inherently expensive even in mass production.
2.2 Advantages and limitations The IRIS system features several key properties which can be viewed as significant advantages over existing technologies: (AI) As seen above, the signal-to-noise ratio is very high; this guarantees high reliability. (A2) The low cost of the components (a laser illuminator similar to those used in CD players and a CCD chip similar to those used in inexpensive surveillance cameras), combined with the simplicity of the concept results in very low cost for the total system: as low as $ 50, which is much lower than any of the currently available technologies. (A3) Because every vehicle on the road has reflective taillights and many have reflective license plates, there is no need for installation of additional infrastructure. (A4) Due to the ability of the retroreflectors in taillights to reflect incident radiation from a wide range of angles back to its source, the IRIS sensor can see vehicles at angles up to 40 degrees on either side; its total field of view is therefore on the order of 80 degrees, which is wider than existing sensors. (A4) There is no need for on-line pointing or focusing; this eliminates any moving parts, hence increasing the robustness of the system while decreasing its cost.
At UCLA we are exploring an innovative approach to ranging called IRIS (Intelligent Ranging with Infrared Sensors), which combines the unique capabilities of infrared (IR) optics with a novel and very simple operating principle. We are building a new class of accurate, robust, and inexpensive ranging sensors which will have more in common with consumer entertainment electronics than with military ranging systems. The IRIS system is composed of a low-power infrared diode laser illuminator, similar to those found in commercial compact disc players, with a relatively wide beam which is pulsed on/off, and a receiver consisting of a charged-coupled device (CCD) image sensor, similar to those used in inexpensive surveillance cameras, behind a narrow bandpass filter. As shown in Figure I, the system records two images of the preceding vehicle in rapid succession: the first one with the illuminating laser turned off, and the second one with the laser on. Due to the micrometer wavelength of the illuminator, almost all surfaces
80
Reduced exposure and laser turned off
Regular exposure
, Reduced exposure and laser turned on
Difference between images with laser on and off Fig. 1. IRIS operating principle: the original scene is reduced to a three-spot pattern which is clearly distinguishable above the clutter.
the next, the IRIS sensor which we have described thus far, can not yield accurate absolute distance measurements when used alone in a general highway environment.
(AS) Ambient light conditions are irrelevant to the
sensing process. (A6) Mechanical vibrations generated by the engine or the road do not affect the sensing process, hence no active vibration cancellation is required.
However, this problem can be eliminated through a straightforward modification of the sensor configuration: instead of using one CCD receiver, one can use two which are mounted at a known lateral distance from each other on the host vehicle. To distinguish between these two configurations, we will denote them as IRIS-I and IRIS-2 depending on whether one or two receivers are used.
Thanks to these features, the IRIS sensors have immediate applications in autonomous ranging for ACC systems currently under development by several automobile manufacturers, as well as the potential for significant long-term impact in AHS deployment, through their use for autonomous and cooperative ranging and even for vehicle-to-vehicle communication. In order to appreciate the full potential of the IRIS system, it is important to understand what its capabilities and limitations are:
The IRIS-2 sensor uses the distance between its receivers as the baseline, and then measures the apparent positions of the same object (such as left or right taillight) on each of its two receivers to implement its triangulation scheme. Thus, IRIS-2 can compute absolute distance at the additional cost of a second receiver and a more elaborate vehicle installation procedure.
Absolute distance. In order to compute absolute distance and relative velocity to the preceding vehicle, the IRIS sensor needs to know a baseline lateral distance, such as the distance between the taillights of that vehicle. This is due to the fact that IRIS uses triangulation schemes, so at least one of the sides of the triangle needs to be of known length. Since the distance between taillights varies from one vehicle to
Although IRIS-I can not measure absolute distance, it can easily lock onto a vehicle in front and track changes in the distance with high reliability, since these changes are measured as changes in the apparent separation of the taillights on the single CCD receiver.
81
ignore the system. Th avoid that, such systems usually set a fairly high return threshold above which an alarm will be triggered. As a result, they often misjudge lowlevel returns from small or low-profile vehicles and motorcycles as false and do not activate their alarm until these targets are much too close for effective action to be taken.
Close following. An additional important difference between IRIS-I and IRIS-2 is in their ability to track vehicles at very small distances. IRIS-I must be able to see both taillights of the preceding vehicle; even with its 80-degree field of view, it may lose one of the taillights at distances smaller than 2 meters. IRIS-2, on the other hand, can maintain tracking at distances as low as 0.5 meter by appropriately adjusting the orientation of one of its receivers.
The IRIS sensor can significantly mitigate this problem when used as a very-low-cost add-on to such systems. By combining the radar returns with the virtually foolproof detection of IRIS, such a system would eliminate the threshold problem for vehicles and other objects with reflective surfaces in most weather conditions, and it would still be able to recognize nonreflective objects and operate in severe weather, albeit in somewhat degraded mode.
Reflective surfaces. Due to its operating principle of subtraction, IRIS cannot detect any objects which do not have reflective surfaces. On the other hand, it can very easily and reliably detect any reflective object. such as vehicle taillights, street signs. overhead signs, and even reflective lane markers. Adverse weather. Like any other sensor operating in visible or near-visible wavelengths, IRIS is not able to operate reliably in very heavy rain, fog, or snow. However, due to the fact that it relies on the subtraction of two images and not on time-of-flight measurements, the operating weather threshold of IRIS is significantly higher than that of other infrared or vision-based sensors.
Adaptive cruise control. Since IRIS-I can only measure changes in the distance from an arbitrary preceding vehicle, it can be used as as a stand-alone ranging sensor for adaptive cruise control systems which use a driver-initialized spacing policy. In other words. the driver is responsible for first achieving the desired spacing and then activating the system; this way, IRIS1 can be used to maintain that selected distance, even though it does not know its exact value.
Lateral motion. When IRIS is locked on a vehicle, it monitors the apparent position and size of its taillights and license plate on the CCD chip. Hence, it can quickly detect and measure the lateral motion as well as the turning rate of the vehicle ahead.
It is well known that constant spacing results in less than ideal performance for adaptive cruise control, and that time-dependent spacing yields better response. This spacing policy, usually called time headway, expresses the desired spacing from the preceding vehicle in seconds and compares it to the current spacing. which is computed as the absolute distance divided by the current speed of the host vehicle. Hence, in order to be able to implement this spacing policy, the controller needs to know the current speed of its own host vehicle (always available) as well as the absolute distance to the vehicle in front, which the IRIS-I sensor can not provide. Nevertheless, the above principle ofdriver initialization applies to time headway as well. In particular, it can be shown that the time headway can be computed up to an unknown constant from the apparent separation of the taillights and the speed of the host vehicle. Hence, once the driver achieves the desired headway and activates the system, it is possible to maintain this headway by controlling the apparent separation in response to changes in the vehicle speed; the unknown constant is not required for this process.
Multiple targets. Because each vehicle in the wide field of view of an IRIS sensor appears as a separate triangular pattern, it is very easy to track multiple targets without being confused by their separate returns. lAne entry and departure. Due to the high reliability of its measurements and the robustness of its operating principle, the IRIS sensor is virtually immune to false returns. Hence, when another vehicle enters the lane in front or when the vehicle that was being tracked leaves that lane, the IRIS sensor immediately knows about it due to its ability to track lateral motion within its wide field of view.
2.3 Potential uses Using the above list of advantages and disadvantages, it is straightforward to conclude that IRIS will be useful in the following ITS applications:
On the other hand, driver initialization is not as userfriendly as the ability to "dial up" a desired distance or headway. Consider, for example, the case where the preceding vehicle leaves the lane and a new vehicle must be tracked: in an IRIS-I system, the driver has to repeat the initialization procedure each time. Clearly, it would be preferable to use an IRIS-2 system, which can close in on the new target without any additional input from the driver. Furthermore, IRIS-2 can be used
Collision warning and avoidance. Clearly, IRIS can not be used as a stand-alone sensor for collision warning and avoidance, since those systems have to be able to detect non-reflective obstacles like pedestrians, trees, and highway barriers, and must also be capable of operating in very low visibility conditions such as dense fog, heavy rain and snow. For these applications it seems radar sensors are necessary. However, radar suffers from shadow returns, which can set off frequent false alarms and prompt the driver to turn off or 82
surfaces, it can be used as a sensor for road curvature. This potential is further enhanced by the ability of IRIS to reliably monitor the movement of several vehicles in its wide field of view, even of vehicles which are several hundred meters away; after all, tracking the motion of preceding vehicles is commonly used by human drivers for predicting road curvature.
to implement mixed spacing policies, which combine constant spacing with an added time headway. It is worth noting that there are several ways of augmenting the capabilities of IRIS-I in order to obtain accurate absolute distance measurements and eliminate the need for driver initialization: (i) Fixed reflective patterns: One can require that all vehicles to be tracked be equipped with three reflective patches (like small bumper stickers) in a fixeddistance pattern. This low-eost solution is easily implementable in fleet settings (truck fleet operators, railroad cars), and it offers the additional benefit of being able to distinguish fleet vehicles from other vehicles on the road through the use of a distinctive reflective pattern. (ii) Additional sensors: If the vehicle already has another sensor installed (microwave radar, laser radar, vision) which has the capability of measuring absolute distance, then IRIS-I may be used as a very-low-eost add-on for enhancing reliability and improving performance.
As it is evident from the above list, IRIS-2 has sig-
nificant potential as a low-eost stand-alone sensor. As for IRIS-I, its most important use may prove to be as a very-Iow-cost add-on to be combined with another sensor such as radar or machine vision. Such combinations will result in enhanced reliability and a larger region of operation, and will be usable in a wide range of ITS applications. 3. NONLINEAR DESIGN OF LONGITUDINAL CONTROLLERS One of the most critical obstacles in the automated operation of commercial heavy vehicles (CHVs) is the presence of significant delays in the fuel and brake actuators. These delays are especially important in longitudinal control of vehicle platoons which do not employ intervehicle communication, because their effect becomes cumulative as it propagates upstream, resulting in considerably degraded performance. At UCLA we have been able to design autonomous controllers which, in the presence of large delays, recover the good performance achieved when the delays are negligible. We present two different approaches which are tailored to different performance requirements and computational resources. A backsteppingbased nonlinear scheme with prediction almost recovers the original "delay-free" performance at the cost of additional controller complexity, while a simpler PID-based nonlinear scheme yields nearly as good performance.
Vehicle following/electronic towbar. In the second stage of vehicle automation, when automated steering is added to automated throttle and brake operation, the control objective becomes more ambitious: Instead of merely alerting the driver to the presence of a vehicle in front or maintaining a desired spacing from it. one is now interested in actually following this vehicle through turns. Since IRIS can easily measure the lateral motion of the preceding vehicle, it can be used as a stand-alone sensor for automated steering. However, in that case the driver has to be certain that the vehicle in front can be trusted to follow the desired path. Therefore, this use of IRIS is very well suited for an electronic towbar scenario, where a commercial truck fleet operator foons an electronic train of several trucks, with the first one driven manually by a human driver and the others following in fully automated driverless mode. In a general highway setting, however, it may be preferable to use a system which follows the vehicle in front only as long as that vehicle remains in the current lane. A lane detection capability is then needed; this is discussed in the next paragraph.
The motivation for this dual approach comes from the fact that we want our controllers to be able to work with new as well as existing trucks and trailers. CHV manufacturers are beginning to equip their vehicles with brake-by-wire systems, commonly referred to as Electronic Braking Systems (EBS), which significantly reduce brake actuator delays in order to meet ever-stricter government regulations on braking distances. While these developments justify our efforts on controller design for vehicles with very small actuator delays, their widespread implementation is yet to come. Furthermore, one has to remember that EBS is even farther away when it comes to trailer brakes, where the largest delays occur. And even if we assume that all future trailers will be equipped with EBS, controller design must still allow for large delays: Since tractor/trailer combinations are mixed and matched, a tractor modified for automated operation must also be able to pull a trailer without EBS.
Lane departure warning and road curvature prediction. The ability to measure the position of the vehicle in the current lane, as well as the curvature of the lane and the road ahead, is critical for applications such as lane departure warning and avoidance. Again, since IRIS can only detect reflective surfaces, it can not be used as a stand-alone sensor for such an application in a general road setting. However, most roads and nearly all highways have some reflective surfaces to aid drivers in predicting road curvature at night. These surfaces include reflective lane separators (bots-dots) in the US, reflective paint for lane marking and roadside posts in Europe, and even roadside and overhead signs. Since IRIS can detect not only the presence but also the shape and the relative velocity of such
In light of these short-term objectives; it becomes imperative to design controllers which will require only
83
80 :
h: Xr :
8d
= 80 + hVf:
VI: Vf: Vr
= tJj -
6=
Xr -
Vf : 8d:
minimum distance between vehicles time headway vehicle separation desired vehicle separation velocity of leading vehicle velocity of following vehicle relative vehicle velocity separation error
Fig. 3. Variable time headway h
= sat(ha -
ChVr).
Fig. 2. Parameters for vehicle following. minimal modifications to vehicles currently in operation and production. Our initial controller designs were based on the fact that for the purpose of AHS participation, CHVs will be equipped with actuators which feature considerably reduced delays. In simulations where delays were assumed to be small, our adaptive nonlinear controllers and nonlinear spacing policies demonstrated robust behavior in demanding merge-and-brake inter-platoon maneuvers, in addition to the objective for which they were designed: maintaining small intra-platoon spacing errors. However, the nonlinear spacing policies which have proven so beneficial in vehicles with negligible actuator delays are not able to cope with the effects of large delays. Accounting for realistic air brake response delays has proven to be a formidable challenge for longitudinal control design in the vehicle following scenario.
--_~- -
- - ------ -
- - - - - - - - - -.",-----
Fig. 4. Variable k choice. ing policies which yield small separation errors without increased intervehicle spacing under autonomous vehicle operation. First, we introduced the notion of variable time headway: instead of being fixed, the headway varies with the relative speed Vr between adjacent vehicles, as shown in Fig. 3.
3.1 Small delays (with EBS)
While we always use a detailed nonlinear vehicle model for simulations, our controller design for the small-delay case is based on a simplified first-order representation of a truck. Our goal was to introduce only as much controller complexity as was necessary in order to meet the perfonnance requirements, and to justify any increase in complexity with a corresponding perfonnance improvement. Hence, these controllers are robust enough to deal with the large discrepancies between the simple model used for design and the detailed one used for simulation.
The intuition behind this modification is as follows: Suppose that a vehicle wants to maintain a time headway of ha from the preceding vehicle, when both of them are traveling at the same speed. If the relative speed between the two vehicles is positive, that is, if the preceding vehicle is moving faster, then it is safe to reduce this headway, while if the preceding vehicle is moving slower then it would be advisable to increase the headway. The effect of introducing the variable time headway is quite dramatic: it results in an impressive reduction of errors and a considerably smoother control activity without any increase in steady-state intervehicle spacing.
The parameters relevant to any two adjacent vehicles in a platoon are illustrated in Fig. 2. In the platoon scenario, the controller has to regulate to zero both the relative velocity V r and the separation erroro.
Another modification is the introduction of a variable separation error gain k which is a function of the separation error 0:
Using existing control schemes, the perfonnance of a platoon with several vehicles is safe and acceptable only when time headway is used. However, when this time headway is constant, it has to be significantly larger than for passenger cars in order to guarantee good CHV platoon perfonnance. In our research, we focused on the development of new nonlinear spac-
k
= Ck + (ko -
2
ck)e-/76
,
(1)
where 0 < Ck < ko and q ~ 0 are design constants. The resulting fonn of the separation error gain is shown in Fig.4. 84
The intuition here is that when the separation error gain k is constant, the controller will tty to reduce a very large spacing error 6 through a very large relative velocity Vr of opposite sign, in order to meet the desired control objective, which is Vr + k6 = O. Hence, if a vehicle falls far behind the preceding vehicle, its controller will react aggressively by accelerating to a very high speed. This undesirable behavior can be corrected by decreasing the gain k as 6 becomes large and positive, making sure that it remains above some reasonable positive lower bound; this results in a smooth reduction of large spacing errors. The expression given in (I) does that, but also has another feature which at first glance may seem counter-intuitive: The gain k is reduced even when 6 becomes negative. This feature is included due to the low actuation-to-weight ratio ofCHVs, which severely limits the accelerations and decelerations they are capable of achieving. In autonomous operation, where each vehicle relies only on its own measurements of relative speed and distance from the preceding vehicle, aggressive control actions are amplified as they propagate upstream. Hence, during a sudden braking maneuver in a CHV platoon, only the first few vehicles will be able to achieve the necessary decelerations; the controllers of the next vehicles will quickly saturate, and collisions may occur. Reducing the gain k for negative 6 makes the reaction of the first few vehicles less aggressive and allows the remaining vehicles to follow safely, thus endows each vehicle's controller with a "group conscience", which sacrifices the individual performance of the first few vehicles in order to improve the overall behavior of the platoon.
20,...---------,
. . - ..... (m)
5
0
eo
-6
0
20
40
eo
_--.(m)
8
I
o
..............
-...., .....
5
I
-2
3
-3
2
~
5
o
20
eo
00
20
40
eo
Fig. 5. No delay, original controller with nonlinear spacing policy. _ _ elTOf(m)
20,...---------,
5,...--.:..........------:rrn
15 10
Ol~-ol'!'!o~ -1
-2 -3 ~
-50'---:02O-:---~40-:-----:!eo
To illustrate our control design, we use simulations of a platoon comprising seven (7) tractor-semitrailer combination vehicles. Both fuel and brake actuator have a pure time delay T = 0.2 s each. The platoon starts out at an initial speed of 12 mls. At t = 10 s the platoon leader is given a command to accelerate at 0.2 mls 2 for 10 s. Then at t = 35 s a command for deceleration at 3 mls 2 is issued for 3 s. The minimum desired separation between vehicles is 80 = 3 m. This demanding scenario is representative of the difficulties the system might have maintaining stable platoon behavior when ttying to meet a challenging acceleration/deceleration objective. In all our simulation plots, different vehicles are represented by lines of different thickness: Vehicle 1 is shown with a thick solid line, while lines corresponding to the following vehicles become thinner as the vehicle's number in the platoon increases. The desired velocity profile is given in a dash-dotted line.
eo
Fig. 6. Actuator delay T = 0.2 s, original PIQ controller with nonlinear spacing policy.
3.2 lArge delays (without EBS) When we also include significant actuator delays in the simulation model, we remove some of this discrepancy by using a second-order model which includes the actuator dynamics. Starting from our original controller and using a backstepping procedure, we derive a new control law which demonstrates significantly improved performance in the presence of large actuator delays. However, in the presence of significant delays, the same controller with variable h and k cannot yield acceptable performance because its gains have to be reduced in order to maintain stability. Multiple crashes are observed in the "vehicle separation" plot of Fig. 6 due to the abrupt deceleration maneuver commanded from t 35 s to t 38 s.
The original PIQ controller with both nonlinear spacing policies, variable time headway h = 0.1 - 0.2vr s and variable separation error gain k = 0.1 + (1 0.I)e- O. 162 yields good performance in the absence of delays as seen in Fig. 5.
=
=
Therefore, we designed a nonlinear PID-Iike controller, using the same nonlinear spacing policy along with a filtered version ("dirty derivative") of the relative velocity, which replaces the unavailable relative acceleration measurement. The resulting performance 85
wNcIe wlocIly (.....)
~_(m)
2O,-- ------ -:.---- ,
°0'---2 O---40 -----'eo
6,---------,
-60~-....,2O:':---40:-:-----:!eo _--..(m)
8,----------, 5
OL---"" "",,,-,-1
-2
3
-3
2
5,....- ----,- -----,
°0l---2 -0--40 -----leo
-6'-----'--------' o 20 40
.....
1
i
_ ..,..,.ton
,
h: k:
p:
q:
~."",(m)
20,---------.,
o
Our findings are summarized through a graphical qualitative compar ison of the control schemes. This comparison, shown in Fig. 9, not only summarizes the results presented above, but also allows designers to better negotiate the trade-offs between platoon performance, control smoothness, robustness, and controller complexity in the choice of a scheme which best fits the needs of a particular implementation.
b:
Fig. 7. Actuator delay T = 0.2 s, PID controller with nonlinear spacing policy.
1
comparison
d:
°0l---2 0:'---4 0--...J 80
-wIodt y(.....)
Qualitativ~
circle:
--4
-50~---:2O-:-----:40-:----:'eo
3.3
(m)
8,-----,-------, 5
A
-2
3
-3
2
--4
-6
0
20
40
eo
Fig. 8. Actuator delay T = 0.2 s, backstepping controller with discrete predict or and nonlinear spacing policy.
Fig. 9. Qualitative comparison of control schemes. ACKNOWLEDGMENTIDISCLAIMER This work is supported in part by the U.S. Department of Transportation through the ITS-IDEA Program of the Transportation Research Board, National Academy of Sciences, under Contract ITS-61, and in part by the California Department of Transportation (CalTrans) through the California PATH Program under MOU 314. The contents of this paper reflect the views of the authors who are responsible for the facts and the accuracy of the data presented herein. The contents do not necessarily reflect the official views or policies of the U.S. Government or of the State of California. This paper does not constitute a standard, specification or regulation.
is very good, as seen in Fig. 7, and it is achieved with only a slight increase in controller complexity. Moreover, the PID scheme demonstrates reasonable robustness with respect to the value of the delay. We should not that regular PID controllers which do not utilize our nonlinear spacing policies were incapable of achieving acceptable perform ance in our simulations. Finally, we designed a more advanced nonlinear controller using the so-called "backst epping" technique, which utilizes a second-order vehicle model, and combined it with a discrete-time predict or which attempts to overcome the effect of the delay by extrapolating from current and past values of the measured quantities. This is the most comple x control algorithm of the group, but it yields dramatic improvements. As seen in Fig. 8, it nearly recaptures the delay-free performance of Fig. 5.
FURTHER INFORMATION For additional information, please visit the Website of this project:
http://ansl.ee.ucla.eduliris
86
INTELLIGENT SENSORS AND CONTROL FOR COMMERCIAL VEmCLE AUTOMATION
loannis Kanellakopoulos
PbyUis Nelson Osc:ar Stafsudd
UCLA Electrical Engineering, Los Angeles, CA 90095-1594 Abstract: One of the potentially most prominent applications of Intelligent Transportation Systems technology is the partial or full automation of commercial vehicles such as trucks and buses for freight and passenger transport. The design issues which are particular to these vehicles have largely been ignored thus far, especially issues related to sensor and actuator limitations. Here we address the longitudinal control problem for commercial heavy vehicles, and we present a new class of ranging sensors and a new class of fueVbrake controllers which eliminate or bypass many of these limitations, and can be used in all stages of ITS deployment. Copyright© 1998 IFAC Keywords: Commercial vehicles, NHS, Ranging sensors, Intelligent systems, Longitudinal control, Nonlinear control
1. INfRODUCTION
manding tasks than autonomous systems, such as coordinated driving in a group, but their time to commercialization is likely to be longer. 3. Automated highway systems add information obtained from the roadway infrastructure, such as messages regarding traffic conditions and road geometry, and lateral information from magnetic nails or reflective guardrails installed on the highway. Such systems can perform even more demanding tasks, like fully automated driving in a platoon, but must face many more obstacles (standardization, liability issues, public acceptance) on their way to implementation.
Intelligent Transportation Systems (ITS) technology is progressing at an ever-increasing rate, with exciting developments in all fronts, from driver information and assistance systems to Automated Highway Systems (AHS). This presentation discusses potential scenarios with varying levels of automation for commercial vehicles, and introduces a new sensor technology called IRIS (Intelligent Ranging with Infrared Sensors) and new nonlinear control algorithms which can be combined with existing sensors and actuators to produce economically feasible automation solutions.
Commercial vehicles, in particular, will reap significant benefits from all stages of automation. Collision warning systems increase safety and reduce accidents, while adaptive cruise control enhances driver comfort and reduces fuel consumption and emissions. Fleet operators can further reduce their costs using cooperative scenarios like the electronic towbar, in which one manually driven vehicle is followed by two or three driverless automated vehicles. Finally, in automated highway systems, fully automated vehicles will be able to carry freight and passengers with significantly enhanced safety, increased fuel efficiency, and much more predictable travel times.
One of the main points of contention in the ITS community is the level of vehicle-to-vehicle and vehicleto-roadway cooperation that can be assumed in current research. In that respect, systems currently in various stages of research and development can be classified into three categories: 1. Autonomous systems depend only on information obtained by the sensors located on the vehicle itself, usually relative distance and velocity to stationary objects and moving vehicles. They are therefore implementable in the immediate future, and in fact have started to appear as commercial products (collision warning, adaptive cruise control).
Two critical automation components in any of these scenarios, especially for commercial vehicles, are the sensors and the actuators. Existing sensor technologies can only provide the reliability required for these
2. Cooperative systems add information transmitted by neighboring vehicles, usually acceleration and steering inputs. Hence, they can perform more de-
79
of the preceding vehicle and the surrounding objects scatter the incident laser beam and reflect only a negligible fraction back to the receiver to be recorded on the second (laser on) image. There are, however, some surfaces which reflect most of the incident laser beam back to its source: these are the passive reflective patches molded into the taillights of all modern vehicles, as well the reflective paint used in the license plates of many states, which act as efficient retroreflectors in the illuminator's wavelength. Subtraction of the first (laser off) image from the second (laser on) one yields a clear sharp image of these reflective surfaces against a flat background. The distance to the preceding vehicle is then easily computed from the resulting clear pattern via standard triangulation schemes.
applications by increasing the cost to levels beyond those acceptable to the potential customers. Actuators, on the other hand, have built-in saturations and delays which significantly complicate the control task. These issues are addressed by our current research, in which we are developing (1) very inexpensive and highly accurate ranging sensors that can be used as stand-alone sensors for adaptive cruise control and vehicle following, and also in combination with radar or vision for improved collision warning, lane tracking, as well as driving in platoons, and (2) novel nonlinear control algorithms which can deal with the severe actuator saturations and delays present in commercial vehicles to yield good performance in both autonomous and cooperative scenarios. The reason for focusing on these projects is that we are interested in results that will be useful in all stages of automation and will potentially impact product development ranging from today's collision warning and adaptive cruise control to the future's fully automated AHS vehicles.
It is particularly important to note that even in this simple example the clutter is essentially nonexistent. This is due to the combined effect of (1) the narrow bandpass filter, which significantly reduces the contribution of broadband sources such as the sun or artificial lights without affecting the laser returns, and (2) the image subtraction process, which eliminates all returns except those produced by laser illumination.
2. IRIS: INTELLIGENT RANGING USING INFRARED SENSORS
With the modulated illuminator and synchronous detection of the IRIS system, the retroreflectors and license plate are the dominant features in the subtracted image. The IRIS computational requirements are thus significantly lower than those of a conventional approach, which would use two-dimensional image processing techniques to extract the position of the vehicle from the cluttered scene of the original image.
2.1 Operating principle Reliable ranging between adjacent vehicles is one of the key enabling technologies for ITS. However, the dynamically changing random environment of highway traffic makes it very difficult for existing technologies to achieve the necessary accuracy and reliability. In the case of radar and sonar, this is due to their susceptibility to secondary reflections, while vision-based systems require massive computations. Furthermore, most of these technologies are based on component parts which are inherently expensive even in mass production.
2.2 Advantages and limitations The IRIS system features several key properties which can be viewed as significant advantages over existing technologies: (AI) As seen above, the signal-to-noise ratio is very high; this guarantees high reliability. (A2) The low cost of the components (a laser illuminator similar to those used in CD players and a CCD chip similar to those used in inexpensive surveillance cameras), combined with the simplicity of the concept results in very low cost for the total system: as low as $ 50, which is much lower than any of the currently available technologies. (A3) Because every vehicle on the road has reflective taillights and many have reflective license plates, there is no need for installation of additional infrastructure. (A4) Due to the ability of the retroreflectors in taillights to reflect incident radiation from a wide range of angles back to its source, the IRIS sensor can see vehicles at angles up to 40 degrees on either side; its total field of view is therefore on the order of 80 degrees, which is wider than existing sensors. (A4) There is no need for on-line pointing or focusing; this eliminates any moving parts, hence increasing the robustness of the system while decreasing its cost.
At UCLA we are exploring an innovative approach to ranging called IRIS (Intelligent Ranging with Infrared Sensors), which combines the unique capabilities of infrared (IR) optics with a novel and very simple operating principle. We are building a new class of accurate, robust, and inexpensive ranging sensors which will have more in common with consumer entertainment electronics than with military ranging systems. The IRIS system is composed of a low-power infrared diode laser illuminator, similar to those found in commercial compact disc players, with a relatively wide beam which is pulsed on/off, and a receiver consisting of a charged-coupled device (CCD) image sensor, similar to those used in inexpensive surveillance cameras, behind a narrow bandpass filter. As shown in Figure I, the system records two images of the preceding vehicle in rapid succession: the first one with the illuminating laser turned off, and the second one with the laser on. Due to the micrometer wavelength of the illuminator, almost all surfaces
80
Reduced exposure and laser turned off
Regular exposure
, Reduced exposure and laser turned on
Difference between images with laser on and off Fig. 1. IRIS operating principle: the original scene is reduced to a three-spot pattern which is clearly distinguishable above the clutter.
the next, the IRIS sensor which we have described thus far, can not yield accurate absolute distance measurements when used alone in a general highway environment.
(AS) Ambient light conditions are irrelevant to the
sensing process. (A6) Mechanical vibrations generated by the engine or the road do not affect the sensing process, hence no active vibration cancellation is required.
However, this problem can be eliminated through a straightforward modification of the sensor configuration: instead of using one CCD receiver, one can use two which are mounted at a known lateral distance from each other on the host vehicle. To distinguish between these two configurations, we will denote them as IRIS-I and IRIS-2 depending on whether one or two receivers are used.
Thanks to these features, the IRIS sensors have immediate applications in autonomous ranging for ACC systems currently under development by several automobile manufacturers, as well as the potential for significant long-term impact in AHS deployment, through their use for autonomous and cooperative ranging and even for vehicle-to-vehicle communication. In order to appreciate the full potential of the IRIS system, it is important to understand what its capabilities and limitations are:
The IRIS-2 sensor uses the distance between its receivers as the baseline, and then measures the apparent positions of the same object (such as left or right taillight) on each of its two receivers to implement its triangulation scheme. Thus, IRIS-2 can compute absolute distance at the additional cost of a second receiver and a more elaborate vehicle installation procedure.
Absolute distance. In order to compute absolute distance and relative velocity to the preceding vehicle, the IRIS sensor needs to know a baseline lateral distance, such as the distance between the taillights of that vehicle. This is due to the fact that IRIS uses triangulation schemes, so at least one of the sides of the triangle needs to be of known length. Since the distance between taillights varies from one vehicle to
Although IRIS-I can not measure absolute distance, it can easily lock onto a vehicle in front and track changes in the distance with high reliability, since these changes are measured as changes in the apparent separation of the taillights on the single CCD receiver.
81
ignore the system. Th avoid that, such systems usually set a fairly high return threshold above which an alarm will be triggered. As a result, they often misjudge lowlevel returns from small or low-profile vehicles and motorcycles as false and do not activate their alarm until these targets are much too close for effective action to be taken.
Close following. An additional important difference between IRIS-I and IRIS-2 is in their ability to track vehicles at very small distances. IRIS-I must be able to see both taillights of the preceding vehicle; even with its 80-degree field of view, it may lose one of the taillights at distances smaller than 2 meters. IRIS-2, on the other hand, can maintain tracking at distances as low as 0.5 meter by appropriately adjusting the orientation of one of its receivers.
The IRIS sensor can significantly mitigate this problem when used as a very-low-cost add-on to such systems. By combining the radar returns with the virtually foolproof detection of IRIS, such a system would eliminate the threshold problem for vehicles and other objects with reflective surfaces in most weather conditions, and it would still be able to recognize nonreflective objects and operate in severe weather, albeit in somewhat degraded mode.
Reflective surfaces. Due to its operating principle of subtraction, IRIS cannot detect any objects which do not have reflective surfaces. On the other hand, it can very easily and reliably detect any reflective object. such as vehicle taillights, street signs. overhead signs, and even reflective lane markers. Adverse weather. Like any other sensor operating in visible or near-visible wavelengths, IRIS is not able to operate reliably in very heavy rain, fog, or snow. However, due to the fact that it relies on the subtraction of two images and not on time-of-flight measurements, the operating weather threshold of IRIS is significantly higher than that of other infrared or vision-based sensors.
Adaptive cruise control. Since IRIS-I can only measure changes in the distance from an arbitrary preceding vehicle, it can be used as as a stand-alone ranging sensor for adaptive cruise control systems which use a driver-initialized spacing policy. In other words. the driver is responsible for first achieving the desired spacing and then activating the system; this way, IRIS1 can be used to maintain that selected distance, even though it does not know its exact value.
Lateral motion. When IRIS is locked on a vehicle, it monitors the apparent position and size of its taillights and license plate on the CCD chip. Hence, it can quickly detect and measure the lateral motion as well as the turning rate of the vehicle ahead.
It is well known that constant spacing results in less than ideal performance for adaptive cruise control, and that time-dependent spacing yields better response. This spacing policy, usually called time headway, expresses the desired spacing from the preceding vehicle in seconds and compares it to the current spacing. which is computed as the absolute distance divided by the current speed of the host vehicle. Hence, in order to be able to implement this spacing policy, the controller needs to know the current speed of its own host vehicle (always available) as well as the absolute distance to the vehicle in front, which the IRIS-I sensor can not provide. Nevertheless, the above principle ofdriver initialization applies to time headway as well. In particular, it can be shown that the time headway can be computed up to an unknown constant from the apparent separation of the taillights and the speed of the host vehicle. Hence, once the driver achieves the desired headway and activates the system, it is possible to maintain this headway by controlling the apparent separation in response to changes in the vehicle speed; the unknown constant is not required for this process.
Multiple targets. Because each vehicle in the wide field of view of an IRIS sensor appears as a separate triangular pattern, it is very easy to track multiple targets without being confused by their separate returns. lAne entry and departure. Due to the high reliability of its measurements and the robustness of its operating principle, the IRIS sensor is virtually immune to false returns. Hence, when another vehicle enters the lane in front or when the vehicle that was being tracked leaves that lane, the IRIS sensor immediately knows about it due to its ability to track lateral motion within its wide field of view.
2.3 Potential uses Using the above list of advantages and disadvantages, it is straightforward to conclude that IRIS will be useful in the following ITS applications:
On the other hand, driver initialization is not as userfriendly as the ability to "dial up" a desired distance or headway. Consider, for example, the case where the preceding vehicle leaves the lane and a new vehicle must be tracked: in an IRIS-I system, the driver has to repeat the initialization procedure each time. Clearly, it would be preferable to use an IRIS-2 system, which can close in on the new target without any additional input from the driver. Furthermore, IRIS-2 can be used
Collision warning and avoidance. Clearly, IRIS can not be used as a stand-alone sensor for collision warning and avoidance, since those systems have to be able to detect non-reflective obstacles like pedestrians, trees, and highway barriers, and must also be capable of operating in very low visibility conditions such as dense fog, heavy rain and snow. For these applications it seems radar sensors are necessary. However, radar suffers from shadow returns, which can set off frequent false alarms and prompt the driver to turn off or 82
surfaces, it can be used as a sensor for road curvature. This potential is further enhanced by the ability of IRIS to reliably monitor the movement of several vehicles in its wide field of view, even of vehicles which are several hundred meters away; after all, tracking the motion of preceding vehicles is commonly used by human drivers for predicting road curvature.
to implement mixed spacing policies, which combine constant spacing with an added time headway. It is worth noting that there are several ways of augmenting the capabilities of IRIS-I in order to obtain accurate absolute distance measurements and eliminate the need for driver initialization: (i) Fixed reflective patterns: One can require that all vehicles to be tracked be equipped with three reflective patches (like small bumper stickers) in a fixeddistance pattern. This low-eost solution is easily implementable in fleet settings (truck fleet operators, railroad cars), and it offers the additional benefit of being able to distinguish fleet vehicles from other vehicles on the road through the use of a distinctive reflective pattern. (ii) Additional sensors: If the vehicle already has another sensor installed (microwave radar, laser radar, vision) which has the capability of measuring absolute distance, then IRIS-I may be used as a very-low-eost add-on for enhancing reliability and improving performance.
As it is evident from the above list, IRIS-2 has sig-
nificant potential as a low-eost stand-alone sensor. As for IRIS-I, its most important use may prove to be as a very-Iow-cost add-on to be combined with another sensor such as radar or machine vision. Such combinations will result in enhanced reliability and a larger region of operation, and will be usable in a wide range of ITS applications. 3. NONLINEAR DESIGN OF LONGITUDINAL CONTROLLERS One of the most critical obstacles in the automated operation of commercial heavy vehicles (CHVs) is the presence of significant delays in the fuel and brake actuators. These delays are especially important in longitudinal control of vehicle platoons which do not employ intervehicle communication, because their effect becomes cumulative as it propagates upstream, resulting in considerably degraded performance. At UCLA we have been able to design autonomous controllers which, in the presence of large delays, recover the good performance achieved when the delays are negligible. We present two different approaches which are tailored to different performance requirements and computational resources. A backsteppingbased nonlinear scheme with prediction almost recovers the original "delay-free" performance at the cost of additional controller complexity, while a simpler PID-based nonlinear scheme yields nearly as good performance.
Vehicle following/electronic towbar. In the second stage of vehicle automation, when automated steering is added to automated throttle and brake operation, the control objective becomes more ambitious: Instead of merely alerting the driver to the presence of a vehicle in front or maintaining a desired spacing from it. one is now interested in actually following this vehicle through turns. Since IRIS can easily measure the lateral motion of the preceding vehicle, it can be used as a stand-alone sensor for automated steering. However, in that case the driver has to be certain that the vehicle in front can be trusted to follow the desired path. Therefore, this use of IRIS is very well suited for an electronic towbar scenario, where a commercial truck fleet operator foons an electronic train of several trucks, with the first one driven manually by a human driver and the others following in fully automated driverless mode. In a general highway setting, however, it may be preferable to use a system which follows the vehicle in front only as long as that vehicle remains in the current lane. A lane detection capability is then needed; this is discussed in the next paragraph.
The motivation for this dual approach comes from the fact that we want our controllers to be able to work with new as well as existing trucks and trailers. CHV manufacturers are beginning to equip their vehicles with brake-by-wire systems, commonly referred to as Electronic Braking Systems (EBS), which significantly reduce brake actuator delays in order to meet ever-stricter government regulations on braking distances. While these developments justify our efforts on controller design for vehicles with very small actuator delays, their widespread implementation is yet to come. Furthermore, one has to remember that EBS is even farther away when it comes to trailer brakes, where the largest delays occur. And even if we assume that all future trailers will be equipped with EBS, controller design must still allow for large delays: Since tractor/trailer combinations are mixed and matched, a tractor modified for automated operation must also be able to pull a trailer without EBS.
Lane departure warning and road curvature prediction. The ability to measure the position of the vehicle in the current lane, as well as the curvature of the lane and the road ahead, is critical for applications such as lane departure warning and avoidance. Again, since IRIS can only detect reflective surfaces, it can not be used as a stand-alone sensor for such an application in a general road setting. However, most roads and nearly all highways have some reflective surfaces to aid drivers in predicting road curvature at night. These surfaces include reflective lane separators (bots-dots) in the US, reflective paint for lane marking and roadside posts in Europe, and even roadside and overhead signs. Since IRIS can detect not only the presence but also the shape and the relative velocity of such
In light of these short-term objectives; it becomes imperative to design controllers which will require only
83
80 :
h: Xr :
8d
= 80 + hVf:
VI: Vf: Vr
= tJj -
6=
Xr -
Vf : 8d:
minimum distance between vehicles time headway vehicle separation desired vehicle separation velocity of leading vehicle velocity of following vehicle relative vehicle velocity separation error
Fig. 3. Variable time headway h
= sat(ha -
ChVr).
Fig. 2. Parameters for vehicle following. minimal modifications to vehicles currently in operation and production. Our initial controller designs were based on the fact that for the purpose of AHS participation, CHVs will be equipped with actuators which feature considerably reduced delays. In simulations where delays were assumed to be small, our adaptive nonlinear controllers and nonlinear spacing policies demonstrated robust behavior in demanding merge-and-brake inter-platoon maneuvers, in addition to the objective for which they were designed: maintaining small intra-platoon spacing errors. However, the nonlinear spacing policies which have proven so beneficial in vehicles with negligible actuator delays are not able to cope with the effects of large delays. Accounting for realistic air brake response delays has proven to be a formidable challenge for longitudinal control design in the vehicle following scenario.
--_~- -
- - ------ -
- - - - - - - - - -.",-----
Fig. 4. Variable k choice. ing policies which yield small separation errors without increased intervehicle spacing under autonomous vehicle operation. First, we introduced the notion of variable time headway: instead of being fixed, the headway varies with the relative speed Vr between adjacent vehicles, as shown in Fig. 3.
3.1 Small delays (with EBS)
While we always use a detailed nonlinear vehicle model for simulations, our controller design for the small-delay case is based on a simplified first-order representation of a truck. Our goal was to introduce only as much controller complexity as was necessary in order to meet the perfonnance requirements, and to justify any increase in complexity with a corresponding perfonnance improvement. Hence, these controllers are robust enough to deal with the large discrepancies between the simple model used for design and the detailed one used for simulation.
The intuition behind this modification is as follows: Suppose that a vehicle wants to maintain a time headway of ha from the preceding vehicle, when both of them are traveling at the same speed. If the relative speed between the two vehicles is positive, that is, if the preceding vehicle is moving faster, then it is safe to reduce this headway, while if the preceding vehicle is moving slower then it would be advisable to increase the headway. The effect of introducing the variable time headway is quite dramatic: it results in an impressive reduction of errors and a considerably smoother control activity without any increase in steady-state intervehicle spacing.
The parameters relevant to any two adjacent vehicles in a platoon are illustrated in Fig. 2. In the platoon scenario, the controller has to regulate to zero both the relative velocity V r and the separation erroro.
Another modification is the introduction of a variable separation error gain k which is a function of the separation error 0:
Using existing control schemes, the perfonnance of a platoon with several vehicles is safe and acceptable only when time headway is used. However, when this time headway is constant, it has to be significantly larger than for passenger cars in order to guarantee good CHV platoon perfonnance. In our research, we focused on the development of new nonlinear spac-
k
= Ck + (ko -
2
ck)e-/76
,
(1)
where 0 < Ck < ko and q ~ 0 are design constants. The resulting fonn of the separation error gain is shown in Fig.4. 84
The intuition here is that when the separation error gain k is constant, the controller will tty to reduce a very large spacing error 6 through a very large relative velocity Vr of opposite sign, in order to meet the desired control objective, which is Vr + k6 = O. Hence, if a vehicle falls far behind the preceding vehicle, its controller will react aggressively by accelerating to a very high speed. This undesirable behavior can be corrected by decreasing the gain k as 6 becomes large and positive, making sure that it remains above some reasonable positive lower bound; this results in a smooth reduction of large spacing errors. The expression given in (I) does that, but also has another feature which at first glance may seem counter-intuitive: The gain k is reduced even when 6 becomes negative. This feature is included due to the low actuation-to-weight ratio ofCHVs, which severely limits the accelerations and decelerations they are capable of achieving. In autonomous operation, where each vehicle relies only on its own measurements of relative speed and distance from the preceding vehicle, aggressive control actions are amplified as they propagate upstream. Hence, during a sudden braking maneuver in a CHV platoon, only the first few vehicles will be able to achieve the necessary decelerations; the controllers of the next vehicles will quickly saturate, and collisions may occur. Reducing the gain k for negative 6 makes the reaction of the first few vehicles less aggressive and allows the remaining vehicles to follow safely, thus endows each vehicle's controller with a "group conscience", which sacrifices the individual performance of the first few vehicles in order to improve the overall behavior of the platoon.
20,...---------,
. . - ..... (m)
5
0
eo
-6
0
20
40
eo
_--.(m)
8
I
o
..............
-...., .....
5
I
-2
3
-3
2
~
5
o
20
eo
00
20
40
eo
Fig. 5. No delay, original controller with nonlinear spacing policy. _ _ elTOf(m)
20,...---------,
5,...--.:..........------:rrn
15 10
Ol~-ol'!'!o~ -1
-2 -3 ~
-50'---:02O-:---~40-:-----:!eo
To illustrate our control design, we use simulations of a platoon comprising seven (7) tractor-semitrailer combination vehicles. Both fuel and brake actuator have a pure time delay T = 0.2 s each. The platoon starts out at an initial speed of 12 mls. At t = 10 s the platoon leader is given a command to accelerate at 0.2 mls 2 for 10 s. Then at t = 35 s a command for deceleration at 3 mls 2 is issued for 3 s. The minimum desired separation between vehicles is 80 = 3 m. This demanding scenario is representative of the difficulties the system might have maintaining stable platoon behavior when ttying to meet a challenging acceleration/deceleration objective. In all our simulation plots, different vehicles are represented by lines of different thickness: Vehicle 1 is shown with a thick solid line, while lines corresponding to the following vehicles become thinner as the vehicle's number in the platoon increases. The desired velocity profile is given in a dash-dotted line.
eo
Fig. 6. Actuator delay T = 0.2 s, original PIQ controller with nonlinear spacing policy.
3.2 lArge delays (without EBS) When we also include significant actuator delays in the simulation model, we remove some of this discrepancy by using a second-order model which includes the actuator dynamics. Starting from our original controller and using a backstepping procedure, we derive a new control law which demonstrates significantly improved performance in the presence of large actuator delays. However, in the presence of significant delays, the same controller with variable h and k cannot yield acceptable performance because its gains have to be reduced in order to maintain stability. Multiple crashes are observed in the "vehicle separation" plot of Fig. 6 due to the abrupt deceleration maneuver commanded from t 35 s to t 38 s.
The original PIQ controller with both nonlinear spacing policies, variable time headway h = 0.1 - 0.2vr s and variable separation error gain k = 0.1 + (1 0.I)e- O. 162 yields good performance in the absence of delays as seen in Fig. 5.
=
=
Therefore, we designed a nonlinear PID-Iike controller, using the same nonlinear spacing policy along with a filtered version ("dirty derivative") of the relative velocity, which replaces the unavailable relative acceleration measurement. The resulting performance 85
wNcIe wlocIly (.....)
~_(m)
2O,-- ------ -:.---- ,
°0'---2 O---40 -----'eo
6,---------,
-60~-....,2O:':---40:-:-----:!eo _--..(m)
8,----------, 5
OL---"" "",,,-,-1
-2
3
-3
2
5,....- ----,- -----,
°0l---2 -0--40 -----leo
-6'-----'--------' o 20 40
.....
1
i
_ ..,..,.ton
,
h: k:
p:
q:
~."",(m)
20,---------.,
o
Our findings are summarized through a graphical qualitative compar ison of the control schemes. This comparison, shown in Fig. 9, not only summarizes the results presented above, but also allows designers to better negotiate the trade-offs between platoon performance, control smoothness, robustness, and controller complexity in the choice of a scheme which best fits the needs of a particular implementation.
b:
Fig. 7. Actuator delay T = 0.2 s, PID controller with nonlinear spacing policy.
1
comparison
d:
°0l---2 0:'---4 0--...J 80
-wIodt y(.....)
Qualitativ~
circle:
--4
-50~---:2O-:-----:40-:----:'eo
3.3
(m)
8,-----,-------, 5
A
-2
3
-3
2
--4
-6
0
20
40
eo
Fig. 8. Actuator delay T = 0.2 s, backstepping controller with discrete predict or and nonlinear spacing policy.
Fig. 9. Qualitative comparison of control schemes. ACKNOWLEDGMENTIDISCLAIMER This work is supported in part by the U.S. Department of Transportation through the ITS-IDEA Program of the Transportation Research Board, National Academy of Sciences, under Contract ITS-61, and in part by the California Department of Transportation (CalTrans) through the California PATH Program under MOU 314. The contents of this paper reflect the views of the authors who are responsible for the facts and the accuracy of the data presented herein. The contents do not necessarily reflect the official views or policies of the U.S. Government or of the State of California. This paper does not constitute a standard, specification or regulation.
is very good, as seen in Fig. 7, and it is achieved with only a slight increase in controller complexity. Moreover, the PID scheme demonstrates reasonable robustness with respect to the value of the delay. We should not that regular PID controllers which do not utilize our nonlinear spacing policies were incapable of achieving acceptable perform ance in our simulations. Finally, we designed a more advanced nonlinear controller using the so-called "backst epping" technique, which utilizes a second-order vehicle model, and combined it with a discrete-time predict or which attempts to overcome the effect of the delay by extrapolating from current and past values of the measured quantities. This is the most comple x control algorithm of the group, but it yields dramatic improvements. As seen in Fig. 8, it nearly recaptures the delay-free performance of Fig. 5.
FURTHER INFORMATION For additional information, please visit the Website of this project:
http://ansl.ee.ucla.eduliris
86