Toward Automated Vehicle Control Beyond the Stability Limits: Drifting Along a General Path

Exceeding a vehicle's handling limits can lead to strong input coupling and instability—conditions that overwhelm traditional control architectures with independent lateral/longitudinal control and that explicitly assume open-loop stable sideslip dynamics. Yet, drivers in “drifting” competitions routinely operate in this region of state-space, while achieving precise control over both sideslip and the vehicle's path. Developing controllers for automated drifting could provide great insight into the general problem of fully utilizing the entire state-space and ensuring that the widest possible range of maneuvers is available to an autonomous vehicle, should the need arise. This could be particularly relevant when navigating inclement weather or difficult terrain: work by Acosta et al. [1], for example, suggests that driver assistance systems that can access these regimes greatly outperform traditional stability control on gravel surfaces. The ability to harness these coupled dynamics could also help mitigate cases of excessive limit understeer.

Early examples in the literature demonstrated stabilization of the vehicle states about a single drift equilibrium, both in simulation, by Velenis et al. [2], and experiment, by Hindiyeh et al. [3]. Because the typical set of inputs—steering, drive torque, and brakes—renders the problem underactuated, adding on the objective of tracking a desired path is not straightforward. Recently, several works have experimentally demonstrated some form of this combined task for simple, constant-curvature circles: the approach by Werling et al. [4] tracked sideslip and vehicle course, while that by Goh and Gerdes [5] tracked sideslip and a specified path. However, due to restrictive assumptions in vehicle modeling or controller formulation, these approaches cannot be easily extended to more complex trajectories.

Indeed, these examples, as well as other studies of the drifting phenomenon, have used a large range of vehicle models. Two-state single-track models were used early on by Ono et al. [6] and Voser et al. [7] to discuss the unstable dynamics of drifting. Hindiyeh et al. [3] and Goh and Gerdes [5] both used three-state single-track models with forces as direct inputs to formulate their controllers, while Werling et al. [4] explicitly included steering and throttle delay. And in Ref. [2], a four-wheel model with weight transfer and wheelspeed/differential dynamics was linearized about a single drift equilibrium for a linear quadratic regulator. In general, the appropriate level of model fidelity—and consequently, choice of control inputs—is an open question.

Contrasting with the previous approaches, this paper develops and experimentally demonstrates a controller structure for automated drifting that is designed explicitly to handle complex trajectories. Vehicle sideslip and lateral error from a curvilinear path coordinate system are chosen as objectives, and the desired controller dynamics are first derived without assuming a specific vehicle dynamics model, or that the vehicle is near an equilibrium point.

This results in a physically insightful control law expressed in terms of the vehicle's state derivatives. When drifting, a standard assumption is removed—the orientation of the vehicle's velocity vector no longer follows that of the vehicle body. Although this at first seems daunting, the controller derivation instead reveals an opportunity: the rotation rate of the vehicle's velocity vector can now be used directly to track the path. Then, by yawing the vehicle body faster or slower than its velocity vector, we can simultaneously control the sideslip of the vehicle.

To realize this controller, we need a vehicle model that strikes a balance between the various levels of fidelity, yet captures the essential physical phenomena. A condition that distinguishes drifting from normal driving is that the rear tires are continuously fully saturated. Indeed, this is generally an important limiting condition to understand, but is a quick transient event in most aggressive driving situations; sustained drifting allows us to explore this at length.

Analyzing the implications of these saturated-tire dynamics motivates the use of an inner wheelspeed control loop to exploit the force–slip relationship. This then enables the treatment of the rear tire forces as direct inputs in the nonlinear model inversion of a three-state single-track model in the outer loop. This physically motivated, cascaded approach allows us to accurately achieve the controller's required state derivatives over the broad range of conditions encountered in a general trajectory.

Experiments on the full-scale MARTY test vehicle (Fig. 1) confirm the effectiveness of the controller in tracking a path with curvature that varies from 1/7 to 1∕20 m, while speed ranges from 25 to 45 km/h, and sideslip is held at −40 deg. Comparisons to a test run without the inner wheelspeed control loop then lend empirical insight to the importance of accounting for the rear wheelspeed dynamics.

In this section, we first introduce the equations of motion for the force-based single-track model in a curvilinear coordinate system; the tire force models are then described.

The vehicle, shown schematically in Fig. 2, has three states. The total velocity vector, with magnitude V, is pointed at a sideslip angle β from the vehicle's longitudinal axis. The yaw rate of the vehicle's body is then denoted as r.

The position of the vehicle is tracked using a curvilinear coordinate system relative to the reference path, as commonly used throughout the literature. The lateral error e is the distance between the vehicle CG and the point on the reference path closest to it; s is the distance along the reference path to this point. Together, they locate the vehicle relative to the reference path, as shown in Fig. 2.

We use the subscript _ref to denote reference quantities at s along the reference path. The reference course $ϕref$ is the tangent angle of the reference path, and the course error, $Δϕ=ϕ−ϕref$⁠, is the angular deviation between the vehicle course and the reference course.

where we have used a few simplifying assumptions: $sin Δϕ≈Δϕ$ and $VΔϕ˙≫V˙Δϕ$⁠. This is a reasonable assumption as $Δϕ$ is driven small in closed-loop operation.

where κ_ref is the curvature of the reference path.

where δ is the steering angle, a and b are the distance from the CG to the front and rear axles, respectively, and m is the mass of the vehicle.

where F_z is the normal load on the tire, $Cα$ is the cornering stiffness, α is the slip angle, and μ is the coefficient of friction.

The task of the controller is to track the specified path and sideslip using steering δ and rear drivetrain torque τ as inputs. This motivates choosing the lateral error e and sideslip β as the control objectives.

The overall structure of the controller is shown in Fig. 3. In the first part, desired stable dynamics are imposed on e and β simultaneously. This yields a desired course rate $ϕ˙des$⁠, as well as r_syn, a synthetic yaw rate input to the sideslip dynamics. Closing the loop around r_syn then yields a desired yaw acceleration, $r˙des$⁠.

In the next step of the controller, the nonlinear vehicle dynamics model (5) is inverted to transform the pair of desired state derivatives, $ϕ˙des$ and $r˙des$⁠, into a steering angle δ and a desired thrust angle γ_des. The thrust angle is then mapped to a desired wheelspeed, ω_des. Finally, the loop is closed around ω_des to produce a torque command τ.

Noting the kinematic relation (1), $β˙=ϕ˙−r$⁠, and with $ϕ˙$ set by the above path-tracking requirement, we see immediately that yaw rate, relative to the vehicle course rate, can be used as a synthetic input to track sideslip.

This is similar to the approaches in Refs. [3] and [5], except that, rather than assuming a steady-state approximation for $ϕ˙≈rref$⁠, we are using the desired $ϕ˙des$ instead.

where we have again assumed $V˙≈0$ and $Δϕ$ small. As expected, $r˙des=r˙ref$ when the errors are zero. Although we include the first terms in parenthesis in Eq. (13) for completeness, it is worth noting that they are typically quite small compared to the other terms in $r˙des$⁠.

The force-based nonlinear single-track model, described in Eqs. (5)–(7), is numerically inverted to map the desired pair of state derivatives $(ϕ˙des,r˙des)$ to a steering angle δ and a rear thrust angle γ. Because there are two desired state derivatives and two inputs, the total acceleration $V˙$ is an uncontrolled variable that is set by the choice of $(ϕ˙des,r˙des)$⁠. Thus, the speed of the vehicle is not explicitly regulated. However, the coupling between the lateral and the longitudinal dynamics that occurs at high sideslip stabilizes velocity when operating under the imposed control law.

This linear dependence of $ϕ˙des$ on V suggests that velocity should be stable as long as $∂V˙∂ϕ˙<0$ and $V˙$ has a zero-crossing along $r˙=r˙ref$⁠, or more generally, along contours of constant $r˙$⁠. This is determined by the shape of the surface of possible state derivatives, a representative example of which is shown in Fig. 4.

As long as this separatrix lies below the plane $V˙=0$⁠, constraining our model inversion to this “upper” surface should have stable velocity zero-dynamics. This also has the benefit of making the mapping of (⁠$ϕ˙des,r˙des$⁠) one-to-one, simplifying the model inversion (Fig. 6).

Since the shape of the feasible surface changes with the state, this argument is only useful if these conditions are generally true over a broad range of states. Numerically checking this over the large range $0.2<r<1.25 rad/s, 3<V<30 m/s,and−5<β<−50 deg$⁠, and plotting the resultant state-space volume in Fig. 7, we see that this is indeed valid except for very high yaw rates at very low speeds.

This idea of leveraging—rather than fighting—the coupled velocity dynamics leads to an interesting analogical takeaway: when operating in this open-loop unstable region, there is a shuffling of objectives. When driving normally, steering is used to track a path and throttle to regulate speed, and the yaw dynamics yield a stable sideslip response; while drifting, steering, and throttle are used in coordination to track a path and sideslip, and the coupling between the state dynamics yields a stable velocity response.

For a general trajectory, it is possible for the commanded (⁠$ϕ˙des,r˙des$⁠) pair to fall outside the feasible actuation space. For this initial presentation and testing of the controller structure, we elect to handle these cases in a naive way. The value of $ϕ˙des$ computed in Eq. (9) is saturated to fall within the feasible values of $ϕ˙$⁠, before being used in Eq. (11) to compute $r˙des$⁠. The resultant (⁠$ϕ˙des,r˙des$⁠) pair is then simply projected into the feasible space along the $r˙=r˙des$ line. This strategy roughly prioritizes the unstable yaw/sideslip dynamics over the lateral error dynamics.

This simple mapping assumes that wheelspeed dynamics can be ignored and is supported by experimental validation of their controllers. However, other approaches in the literature have not made the same assumption, leaving the question of its importance as an open one. For example, in Ref. [2], Velenis et al. use wheelspeed as one of the inputs of a LQR controller for a system linearized about a single drift equilibrium. The wheelspeed is related to tire force with the Pajecka coupled-slip model [9], and a backstepping controller closes the loop around wheelspeed using the total rear axle torque. The controller is successfully simulated using carsim software.

In this paper, we hope to directly address this question through both theoretical discussions below and experimental comparisons in Sec. 5.

where R is the radius of the tire, ω is the wheelspeed, and $Vtravel,xr$ and $Vtravel,yr$ refer to the longitudinal and lateral components of the vehicle velocity at the tire, respectively.

It is important to note that this relationship agrees with both the brush [10]² and the simplified Pajecka magic formula tire models [9] at full saturation.

where $Iω$ is the rotational inertia of the wheel-tire-drivetrain system, and τ is the applied torque.

Rather than simply mapping a desired rear lateral/longitudinal force to torque, as in Eq. (16), a low-level controller could instead use torque to regulate wheelspeed to values desired by the force–slip relationship (17). This approach offers several potential advantages.

These are plotted against γ in Fig. 10, where it is clearly seen that $∂Fyr∂(μFzr)|Fxr$ grows large relative to the other sensitivities when $|γ| ≲ 0.75 rad$⁠.

Second, these wheelspin dynamics lead to appreciable delays in force generation. With Eqs. (18) and (17) and the parameters for the MARTY test vehicle shown in Table 1, selected time responses for F_xr generation for a step torque input around a typical drift equilibrium at $V=9.5 m/s, r=0.79 rad/s, and β=40 deg$ are shown in Fig. 11. Also plotted is the aggregate 95% rise time over the range of $Fxr=0.05−0.95μFzr$⁠. We see that not only does the rise time vary by almost an order of magnitude in this range but that the values are quite high throughout. A useful point of comparison is the lateral force rise time for the nonsliding front tires that comes about from the tire relaxation length phenomenon. At the same speed, and using a typical estimate for tire relaxation length of $σ≈R=0.33 m$ [11], we expect a front lateral force 95% rise time of $−(σ/V) log (0.05)=0.10 s$⁠. From Fig. 11, we see that the force rise time for the rear sliding tires around the equilibrium force value $Fxr=0.72μFzr$ is more than five times this value. This large delay relative to front lateral force generation is particularly important for the controller design presented here, which dictates coordinated steering and rear longitudinal/lateral force inputs over a broad range of conditions. Using a low-level controller to regulate wheelspeed would allow us to set the closed-loop force rise time at a value small enough that the forces can be considered as direct inputs to the system, as in Eq. (5). This cascaded approach preserves the tractability and intuition of dealing with the force-based equations, while explicitly accounting for the delays arising from wheelspeed dynamics.

where $ωdes,f$ is the desired wheelspeed after a first-order filter with time constant $tω$⁠, and we have used the desired rear longitudinal force $Fxrdes$ to compute the feedforward torque.

where P_r is the proportion of lateral weight transfer that occurs over the rear axle and h is the CG height.

Experiments on the MARTY testbed showcase the effectiveness of the controller for a reference path with speeds between 25 and 45 km/h, curvature between 1/7 and 1/20 m, and sideslip of up to –40 deg.

MARTY, depicted in Fig. 1, is a modified 1981 DMC DeLorean that has been equipped with computer-controlled electric power steering and two electric motors that drive the left-rear and right-rear wheels independently. Vehicle parameters used in the experiment are shown in Table 1. An Oxford Technical Systems RT4003 dual-antenna integrated RTK-GPS/IMU is used to obtain vehicle state information at 250 Hz. The controller is implemented on a dSpace MicroAutoBoxII (DS1401) computer that also runs at 250 Hz, and interfaces with the steering, throttle, and navigation subsystems.

While the controller is derived for a general drifting path, in the interest of convenience and succinctness, we use an existing simple path generation procedure—namely, the “quasi-equilibrium” strategy proposed in Ref. [5].

The path is composed of a sequence of unstable drifting equilibrium points. First, a desired curvature κ_ref and sideslip β_ref profile as a function of path distance s is selected. Then, for each point on the path, the equations of motion (5) are solved given the conditions $β=βref, ϕ˙=κrefV, and V˙=r˙=0$⁠, yielding the reference path values of $ϕ˙ref and Vref$⁠.

The selected profile (Fig. 12) has curvature between 1/7 and 1/20 m, with reference sideslip of up to $−40 deg$ and results in equilibrium speeds between 25 and 45 km/h. The resultant path is shown in black in Fig. 13.

Simple entry and exit clothoids are appended to the start and end of the desired path, and a basic path-tracking controller, similar to that presented by Kritayakirana and Gerdes [13], is used to guide the vehicle on this path. The drifting controller is activated at s = 57m and de-activated at s = 463 m. The controller and vehicle parameters are shown in Table 1.

The controller with wheelspeed control tracks the reference path well, as manifested in good lateral error and sideslip tracking performance. A video of an experimental run with this controller can be found online.³

The measured path of the vehicle is shown in blue in Fig. 13 and adheres very closely to the reference path (black). Indeed, from the lateral error plotted in Fig. 14, we see that the largest deviation is just –0.36 m; the root-mean-square (RMS) lateral error is 0.18 m.

This path-tracking performance is even more impressive when considering that the vehicle is operating around its large target sideslip of $−40 deg$⁠, as seen in Fig. 15. This target value is large compared to the maximum steering range of $±38 deg$ and puts the vehicle close to the limits of actuation. The biggest deviation in sideslip tracking is $−6.1 deg$⁠, and the RMS sideslip tracking error is $2.4 deg$⁠. The measured yaw rate tracks the synthetic yaw rate quite well, except during the initial transition to a drifting condition. It is interesting to note that velocity stays quite close to the quasi-equilibrium values. This suggests that the closed-loop velocity dynamics are indeed stable, which is expected since the states never leave the volume (Fig. 7) wherein we hypothesize this to be true. Additionally, it indicates that the model and parameters used for path planning have quite accurately predicted the equilibrium values.

Steering and rear thrust angle commands are shown in Fig. 16. This trajectory exercises the nonlinear model inversion over a broad range of states, resulting in a similarly wide range of inputs—this is most clearly exemplified by $∼65 deg$ of steering angle the controller sweeps through during the test. Finally, it is useful to confirm the assumption used in deriving the equations of motion for lateral error in Eq. (3): referring to Fig. 17, we see that experimentally, $VΔϕ˙$ is everywhere much larger than $V˙Δϕ$⁠.

To investigate the impact of the cascaded wheelspeed control structure proposed in Sec. 4.2, we compare these results to a test run with wheelspeed control de-activated. In this case, the commanded total rear axle torque is simply $τcommand=RFxrdes$ as in Eq. (16). Because the left and right wheels are independently controlled and mechanically decoupled, a simple low-level proportional controller is used to “lock” their wheelspeeds together by changing the torque distribution between the left and right wheels; this is akin to a closed-loop version of the feedforward torque distribution used in Ref. [5] for autonomous drifting around a single equilibrium.

The states of the vehicle are plotted in red in Fig. 18, with the previous results shown in blue. Although sideslip is generally stabilized around the reference, we see that there are significant oscillations in both sideslip and yaw rate. This behavior is also seen in the path-tracking performance, shown in red in Fig. 19. Although the reference path is generally tracked, both lateral error and course error are far larger and more oscillatory than the case with wheelspeed control shown in blue. Furthermore, we see very large, high-frequency corrections in both steering angle and rear thrust angle, plotted in red in Fig. 20.

One candidate explanation for this behavior, as discussed in Sec. 4.2, is delay between the desired rear forces and the actual rear forces due to wheelspeed dynamics. With three states and three forces (5), it is straightforward to numerically differentiate the measured signals, invert the equations of motion, and obtain measured values of the forces.

A Lissajous-style plot of the measured F_yr and F_xr against their desired values is shown in Fig. 21. Results for the experiment without wheelspeed control (red) show large looping orbits, indicating a significant amount of phase lag. This effect is especially stark when viewed against the results with wheelspeed control (blue), which fall very close to the ideal 1:1 line (black) over the entire test.

This paper derives a physically insightful controller for autonomous drifting along a general path that is formulated in terms of the vehicle's state derivatives. The rotation of the vehicle's velocity vector is used to track the lateral error in a curvilinear coordinate system. Then, sideslip is simultaneously stabilized by yawing the vehicle body relative to this velocity vector.

The required state derivatives are then mapped to actuator inputs using nonlinear model inversion cascaded with a simple wheelspeed control loop. An examination of the dynamics of the saturated rear tires suggests that this approach provides advantages in terms of control bandwidth and robustness to disturbances, compared to assuming longitudinal force is a direct input.

This choice of lateral error and sideslip as objectives is made all the more interesting by the concomitant relegation of velocity to an uncontrolled variable. By analyzing the zero-dynamics of velocity in this controller framework, we draw an interesting analogical takeaway. When driving within the stability limits, steering is used to track the path and throttle to regulate speed, and the yaw dynamics naturally yield a stable sideslip response. When drifting, however, steering and throttle are used in coordination to track a path and sideslip, and the coupling between the state dynamics yields a stable velocity response.

Experimental results for a quasi-equilibrium trajectory with curvature between 1/7 and 1/20 m, speeds between 25 and 45 km/h, and sideslip of up to $−40 deg$ demonstrate accurate tracking of both sideslip and lateral error. Furthermore, comparing these results to a test run without wheelspeed control highlights the effectiveness of the cascaded approach. This points to an interesting area of exploration for stability control systems on the individual tire level: rather than preventing saturation all together, wheel slip could instead be used to control the thrust angle from a sliding tire, leading to novel capabilities.

The theme of expanding the envelope of operation extends to the more general vehicle level as well. The ability to harness—rather than shy away from—these coupled dynamics could greatly assist in certain cases, including those of excessive limit understeer and driving on loose surfaces. The current approaches to trajectory planning often constrain themselves to regions with open-loop stable yaw dynamics, or stricter still, due to either modeling assumptions or unsuitable behavior by the vehicle and underlying controller. This controller was derived with a relatively simple force-based model not unlike those presently used by some planning architectures. As such, it could quickly enable future research on trajectory planning—beyond the “quasi-equilibrium” approximation presented here—that works in concert with the chassis control layer to confidently venture outside these regions and more fully utilize the state-space.

The authors wish to thank the Revs Program at Stanford, Bridgestone Corporation, Brembo Corporation, and Renovo Motors for supporting this work, as well as our Dynamic Design Lab colleagues for assisting during testing at Thunderhill Raceway Park. The authors also specially thank the late Professor J. Karl Hedrick for his mentorship in life in general, and in vehicle dynamics, dynamic surface control, and zero-dynamics in particular.