Trajectory tracking is a crucial task in high driving automation developments. Thus, proper control methods must be designed to safely follow the desired reference path and speed. Model Predictive Control (MPC) is one of the most popular advanced model-based control techniques for this purpose. MPC approaches use the vehicle and tire models to predict the future behavior of the vehicle and compute the optimum control sequence to be applied. Moreover, this latter calculation is carried out considering explicitly the physical constraints of the system and its actuators [1
A critical design decision in MPC-based methods is the selection of a proper model to predict the behavior of the vehicle, as the controller performance will depend heavily on its accuracy and real-time capability. However, this selection is not a trivial task, as covering a broad range of speeds, even the limits of handling, while maintaining real-time capabilities, represents a current engineering challenge [2
In the literature, five vehicle-modeling techniques can be found: point-mass, geometric, kinematic, dynamic, and multi-body vehicle models [3
]. The point-mass modeling considers the vehicle as a particle and it is commonly used in motion planning [5
]. Even though it considers accelerations, it ignores the turning capacity of the vehicle [4
]. The geometric modeling considers the basic geometry of the vehicle and uses its geometric relationships for path tracking [7
]. Although it offers good robustness in most low-speed maneuvers, it ignores the velocity and forces on the vehicle, which causes poor tracking performance at high speeds and transitional maneuvers [9
]. The kinematic modeling is a simplified representation that, besides geometry, considers the orientation, velocity, and acceleration of the vehicle [10
]. It provides appropriate performance at low speed (less than 5 m/s) when tire deformations are small and slips angles on the wheels can be neglected [11
]. Experimental validations have shown a good performance of model-based controllers at low speeds [12
]. However, when the lateral forces on tires increase (e.g., while turning at high speeds), its accuracy is compromised [14
]. The dynamic modeling is a more complex vehicle representation that, besides geometry and kinematics, considers the internal forces and the inertia of the vehicle, providing accurate results in high-speed applications and extreme handling maneuvers [16
]. Its implementation requires a tire model to estimate the longitudinal and lateral tire forces. For this purpose, a linear tire model is typically used, as it represents a good trade-off between computational efficiency and accuracy [19
]. Finally, multi-body modeling is the most accurate representation of vehicle dynamics which is mainly employed as a virtual test platform for driving automation developments. Its high complexity and low computational efficiency make it difficult to implement this method today for real-time applications, therefore it is barely used for motion planning [22
]. A comparison of these models in the context of this work is summarized in Table 1
In MPC designs, kinematic and dynamic models are commonly used, as they present the best accuracy vs. computational cost ratio. Linear MPC approaches have been used to modify dynamically trajectory planners in the case of unexpected situations [23
]. However, as seen in Table 1
, each model presents strengths in different driving scenarios. Hence, in order to improve trajectory tracking and increase the speed range, recent works have proposed to combine these approaches [24
]. This way, some authors have integrated both vehicle models in parallel to estimate more accurately relevant vehicle dynamics behavior, such as the side-slip angle [25
] or the vehicle’s position [26
]. As the previous technique requires computing both models in parallel and increasing the computational effort, in recent years, the so-called model-blending approach has been proposed by some authors [28
]. In this latter method, a model-switching strategy allows for selecting the most appropriate model depending on the driving scenario, allowing for increasing the validity range of the MPC-based vehicle control approach.
In the aforementioned works, to perform the model-blending technique, two aspects are usually considered: (1) the switching condition and (2) the switching method. The switching condition is the criteria used to select the best model for each driving circumstance. This criterion has to be defined using a variable that allows for optimizing the operative range of each model so that the model-blending effectively improves trajectory tracking. In [28
], the longitudinal speed is used as the switching condition, selecting the kinematic model to compute MPC predictions when the vehicle moves at low speed while using the dynamic model when moving at high-speed. However, as model validity is limited by the tire model saturation, it seems more appropriate to use another variable to perform the switching which is directly related to the tires’ forces, such as lateral acceleration. Regarding the switching mode, previous works have presented two different methods—firstly, a sudden switch strategy that instantly changes between kinematic and dynamic models [28
]; secondly, a progressive switching strategy that performs this change more smoothly using a linear approach [29
]. Nevertheless, there is not a recommended procedure to blend models, besides the trial-and-error
In this work, a novel vehicle model blending procedure based on lateral acceleration for MPC is proposed. A coupled longitudinal and lateral vehicle dynamics is considered in the MPC formulations for trajectory tracking. The cornering stiffness is constantly estimated for a linear tire–road interaction. Lateral motion is constrained avoiding lane departures. The proposed approach presents three main contributions: (1) the lateral acceleration is selected as the switching condition instead of the longitudinal velocity, as the former is directly associated with the lateral forces on the tires and it defines a more consistent validity application range for each model; (2) a procedure proposal is presented to find the proper lateral acceleration switching value that allows for getting the best performance in terms of path tracking error; and (3) a thorough comparison of the proposed approach is carried out with the kinematic and dynamic vehicle models, as well as the longitudinal velocity as switching condition.
The structure of the paper is as follows: Section 2
presents the vehicle modeling and the combination method for kinematic and dynamic models; Section 3
defines the proposed model blending procedure tuning to achieve the best path-tracking in vehicle model blending; Section 4
details the case study and the MPC-based control architecture used to evaluate the model-blending approaches; Section 5
shows the results of the simulation experiments using a passenger bus in a realistic urban environment, with a detailed analysis and comparisons between switching methods; Section 6
closes with the conclusions and future works.
3. Tuning Procedure for Model Blending
As detailed in the Introduction, the blending of vehicle models is based on two aspects: (1) the switching condition, and (2) the switching method.
The switching condition is based on a physical measure usually available on the vehicle’s acquisition,
being the most used [28
]. However, this value is typically defined by the designer by a rule of thumb based on several tests. A clear reference for this value is defined by [11
] as 5 m/s, this being the recommended limit to employ the kinematic vehicle model. Nonetheless, this limit does not apply to all cases. For instance, in straight-line motion, lateral forces can be neglected, and the kinematic vehicle model could be considered valid in this condition even after 5 m/s.
The switching method is defined as how the switching condition occurs. As analyzed in Section 1
, two main approaches have been proposed: a sudden or step
change and a progressive one, in which a linear
blending is proposed. According to [29
], a progressive transition between models offers a better response in the vehicle motion control in contrast to sudden switching conditions. However, the obtainment of
values for this progressive blending is a complex task, as more than one reference for the switching condition is necessary and no more than trial-and-error
methods are defined to achieve it.
In this work, a novel approach is proposed. The lateral acceleration
is considered as the switching condition parameter as it can be directly related to the current lateral forces on tires in any condition. The use of this variable is more consistent with the theoretical assumptions referred by [11
] for the kinematic and dynamic vehicle models.
Based on this switching condition, the procedure proposed in Table 2
selects the best switching value of
for model blending in step
methods. The lateral (
) and angular (
) errors are considered as key metrics.
This procedure will be applied for the case study proposed in the next section, and the results will be detailed in Section 5.1
5. Results and Discussion
The performance evaluation of vehicle models and switching methods employed are detailed in this section. The elements in the control architecture defined in Figure 2
and detailed in Section 4
are implemented in a MATLAB/Simulink setup which is used to perform a simulation-based analysis.
Considering the tuning procedure defined in Section 3
, the parameters to perform the three switching methods introduced in Section 4.3
) are defined and evaluated first. In addition, pure kinematic (kin
) and dynamic (dyn
) methods are considered for comparison.
Three complete laps are simulated in the defined scenario (Figure 4
), the results being recorded and evaluated. Eight values for
are defined from 1.1 m/s to 8.8 m/s, equally spaced at 1.1 m/s for each simulation test. This will allow for studying the influence of
and the lateral acceleration (
) in the lateral motion control for the defined route.
5.1. Tuning Procedure for Model Blending
In this section, the procedure defined in Section 3
is applied to select the best switching value for
for model blending in the step
methods. Note that the speed
method is based on the
as proposed in [11
]. In the latter case, a step
method is applied, in which a kinematic model is used below 5 m/s, and a dynamic model at higher speeds.
The results of the step-by-step procedure are detailed next.
Steps 1 to 5:
Once the planned route for trajectory-tracking of Section 4.4
is defined, the vehicle motion control is executed using kin
vehicle models at several
as described previously. The median is estimated for the absolute values
models in a grid of
. In practice, the median provides a better estimation in contrast to mean values for the cut-off
definition pointed in Steps 6a–c
. Results are processed through a two-dimensional convolution [40
] creating surface plots as depicted in Figure 6
results are depicted in Figure 6
a, in blue and red, respectively. The influence of
is remarkable along several
tested, having a clear limit from kin
surface intersections. These surface intersections help to prove the initial hypothesis which presents
as a more appropriate switching condition than
results are depicted in Figure 6
b, in blue and red, respectively. There is no clear influence of
on the improvement of the path-tracking performance in terms of
, as kin
models seem to have similar behavior. These findings motivate the idea of selecting
as the basis for a switching strategy.
Steps 6a–c (stepblending):
A linear regression is calculated from
as showed in Figure 7
a (i.e., considering all the values of
, associated with a certain
and all related
values). The intersection of
is approximately in 1.5 m/s
, this being a useful cut-off value to define the switching condition to
. This value allows for obtaining the lowest
values for kin
models. As stated previously, if the same procedure is applied to
b), no relevant results can be extracted, as both models have similar performance.
Steps 7a–d (linearblending):
becomes relevant around 1 m/s
as depicted in Figure 6
a. This is the lowest
value that can be extracted from the surface intersection, which is useful for defining the initial condition of a progressive switching between models. In addition, the step
switching (defined at 1.5 m/s
) is considered as the point of symmetry to this initial condition. Therefore, the linear
switching is determined from 1 m/s
to 2 m/s
being centered around the step
switching condition. The switching methods for model blending based on
∈ [0, 1] are presented in Figure 8
5.2. Trajectory-Tracking Response Analysis
Results for three-of-eight simulation tests at constant
have been selected for discussion simplicity (2.2 m/s, 5.5 m/s, and 8.8 m/s). The linear
method has been chosen for Figure 8
b–d as it presents the best performance compared to other methods. The route values (black line) are located at zero values on z
-axis as a reference, and the z
-axis limits correspond to minimum and maximum estimation values of
b shows the
of the bus for the linear
method. Although the
is set as constant, note that the MPC regulates the final speed to avoid lane-departures (e.g.,
= 8.8 m/s) as defined in Section 4.5
. Hence, this is considered as a proper performance.
c shows the
of the bus. Larger values are obtained while turning as the
increases. Important transitions are observed mostly on the roundabout at the right-side due to non-smooth planned curvatures. This transitional behavior is observed in
results independently of the tested
, a phenomenon that is not acquired previously in
d shows the
of the bus, which is calculated by considering the road’s center-lane and the current position at each time step. The transitional effects described in
seem to affect the
response, and, therefore, the path tracking.
The former results demonstrate that the MPC with the linear method provides an appropriate trajectory tracking.
5.3. Cornering Stiffness Estimation Analysis
Results for three-of-eight simulation tests at constant
have been selected for discussion simplicity (1.1 m/s, 4.4 m/s and 8.8 m/s). The dyn
method has been chosen for Figure 8
e–f as it uses the cornering stiffness estimation in the whole range of
. The route values (black line) is located at zero values on z
-axis as a reference, and the z
-axis limits correspond to minimum and maximum estimation values of both
for the dyn
method are presented in Figure 8
e–f, respectively. These values are obtained at continuous turning maneuvers both at the front and rear axles. As expected, the estimations behave bumpily when steer angles are near zero, since they imply the inverse of a near-zero matrix (see Section 4.2
). In practical terms, this results in larger variations of slip angles that are later attenuated by the linear Kalman filter. Consequently, the dyn
method deteriorates the path-tracking at straight driving even if the vehicle is driving at high speed. This supports the main rationale to switch to kin
method in this driving condition, even for
> 5 m/s, which is successfully achieved by the proposed
-based blending approach as opposed to the suggestion made by the existing
5.4. Lateral and Angular Error Analysis
a,b shows the statistical distribution of
for five study conditions related to the five analyzed methods, allowing comparison for their trajectory-tracking performance. The boxes span (blue boxes) cover from 2% to 98% of the data, the whiskers span (black lines) cover from 1% to 99% of the data, the median (red horizontal lines) and mean (
, red plus signs) values as statistical metrics assessment.
In Figure 9
a, it can be seen that, for this low speed test track, the kinematic model (kin
) clearly outperforms the dynamic model (dyn
) as expected. However, blending these two models can produce better results than either one of them individually. Note that this particular track has the most low speed turns towards the left, whereas right-hand turns are mostly high speed. This allows for exemplifying the limitations of the speed
method. Its positive
distribution resembles the kinematic model behavior, while the negative side is much closer to the dynamic model. Since they can be correlated to the left and right-hand turns and thus the predominantly high and predominantly low speeds, it becomes clear that blending based on the speed uses either model in some cases where the other one behaves better (i.e., on the high speed turns, it uses the dynamic approximation even if the lateral forces are low and the kinematic model behaves better).
On the other hand, the hereby introduced switching strategies (linear
), based on lateral acceleration, provide better behavior than the use of either kinematic or dynamic models, or even the aforementioned blending approach based on the speed (speed
). This is achieved by actually switching when the lateral forces are significant, thus properly using the best approach in every condition to reduce errors (see Figure 6
), rather than just avoiding singularities (which is the main motivation for the speed based blending). Since most of the track has low lateral acceleration, most of the error distribution for both
-based methods (linear
) resembles the kinematic model behavior (see the blue boxes in Figure 9
a, associated with the 2%–98% data interval). However, the black whiskers do show a significant improvement in reducing the lateral error corresponding to those cases with either high lateral acceleration and low speed or those of high speed and low lateral acceleration, thus proving the advantage of introducing the lateral acceleration as the blending parameter in place of the currently accepted vehicle velocity.
Furthermore, it is noted that the linear blending is slightly better than the step in terms of the maximum dispersion (black whiskers), though the actual advantage of this technique relates to the computational cost, as will become evident in the discussion below.
b shows that the
behaves very similarly regardless of the implementation of either of the analyzed methods, which fit the results shown in Section 5.1
and endorse the decision of considering
surfaces for the blending procedure.
5.5. Computational Cost Analysis
To demonstrate the real-time capability of the presented approach, computational cost analysis has been carried out. The required time to calculate each control cycle of the proposed MPC controllers with the different blending methods has been evaluated. All controllers were execute at a 10 ms period on a LATITUDE E5570 provided with an Intel Core i7-6600U, CPU 2.60GHz (Santa Clara, California, USA). The results are depicted in Figure 9
c, in which the statistical distribution of the solving time is depicted, following the same representation applied to Figure 9
a–b. It can be seen where the worst-case scenario is for the step
approaches, with mean values of 0.04 ms and maximums of nearly 0.08 ms. On the contrary, the linear
method offers the best time efficiency with a mean value of 0.03 ms and a maximum solving time of 0.07 ms. Hence, results demonstrate that computational cost can be reduced by the use of blended models.
Note that all the referred approaches are based on a nonlinear MPC. In this case study, the previously calculated state and input values are used as a seed for the next iteration. Hence, when sudden or abrupt changes are required, the number of iterations required to solve the MPC problem increases significantly as depicted in Figure 9
d. For instance, this happens when a sudden transition from a kinematic to a dynamic model is carried out in the step
method. In this sense, the linear
method reduces the required computational cost by lowering the number of iterations required to solve the optimization problem in the blending procedure to even slightly better values than the simple kinematic model.
The performance of MPC-based tracking controllers in automated vehicles is highly dependent on the selection of the model, either kinematic or dynamic. The kinematic enables accuracy at low , e.g., driving straight. However, the dynamic provides better overall results when becomes representative, e.g., turning or steering transient maneuvers. To cover a wide operational range, switching or blending from one model to the other has been proposed in the literature. In particular, proposed approaches’ use of the longitudinal velocity as the switching condition, which does not offer the best performance. Moreover, there is a lack of works related to the proper tuning of model-blending.
In this study, the use of the as opposed to the is proposed as the switching condition to blend vehicle models within an MPC-based trajectory tracking control. As tire forces are the critical factor for the validity of the kinematic/dynamic models, the is considered as a variable with direct relation to these forces, allowing for increasing the overall performance of the blended approach. Additionally, a formal step-by-step tuning approach is proposed and detailed for two methods: linear and step.
The presented method is tested in a case study with an electric passenger bus in a virtual urban scenario. Results show that the proposed blending approaches based on provide a relative improvement of 15% in terms of , in contrast to the method based on proposed in the literature. Additionally, it allows for reducing the maximum computational cost in 12% if a linear blending approach is used. Moreover, the validity of the tuning procedure is demonstrated.
Future works will assess the concepts presented in this research on both Hardware-in-the-Loop tests verification and real test platform validations, including the implementation issues related to parameters and variable estimations.