Skip to content
ELECTRONICS ENGINEERINGelectronics-engineering.com
Automotive

ACC 5: planning, following, elastic approach, preboost and stop & go

Short, middle, long and extra long time gaps, stability and string stability proved, kinematic approach with elasticity, overtaking preboost, stop & go stopping law, MPC.

https://www.electronics-engineering.com/en/acc-planning-time-gap-string-stability-approach-preboost-stop-and-go/

No ratings yet: be the first to rate this page.

We have the target: its distance d, its speed v_t, its acceleration a_t. Every 20 ms the planner decides which acceleration to request from the car. A single controller is not enough: following in steady state, approaching a much slower vehicle, preparing an overtake and stopping in a queue are different problems with different solutions. The structure we use is that of many series systems, rule-based: each mode proposes an acceleration, the most cautious one wins, then limits and jerk.

Rule-based planner as a block diagram
The planner: four proposals (set speed, following with the preboost term, kinematic approach, change of time gap for overtaking) go into a minimum block; then saturation to the limits of the standard and a jerk limiter. The output is the request a_{req} to torque control.

The constant time gap policy

The desired distance is d_{des} = d_0 + h\,v: it grows with our speed. The errors the planner uses are

    \[ e_d = d - d_{des}, \qquad \Delta v = v_t - v. \]

Why this and not a fixed distance? With a fixed distance a string of cars with ACC amplifies disturbances along the string unless the vehicles communicate; with a time gap this can be avoided, as we prove shortly. The driver chooses h among four values.

The four time gap settings
Left, the desired distance as a function of speed for short (1.0 s), middle (1.4 s), long (1.8 s) and extra long (2.2 s), with the minimum of the standard (0.8 s, dotted). At 130 km/h they range from 40 to 83 m. Right, following a car at 100 km/h, the driver switches from long to extra long (5 s) and then, with two presses, to short (25 s): h (violet) changes by 0.25 s per second and the distance follows it without jerks.

The change of h is deliberately slow: a jump from 2.2 to 1.0 s would mean a distance error of 33 m in an instant, and the controller would accelerate all at once.

Steady-state following

The basic mode is a proportional-derivative controller on the distance error, with the target acceleration fed forward:

    \[ a_f = k_d\, e_d + k_v\, \Delta v + k_a\, a_t, \qquad k_d = 0.15\ \mathrm{s^{-2}},\ k_v = 0.6\ \mathrm{s^{-1}},\ k_a = 0.3. \]

Stability. Let us assume for now that the car achieves the requested acceleration immediately (\dot v = a_f) and that the target drives at constant speed. Then \dot e_d = \Delta v - h\,\dot v and \dot{\Delta v} = -\dot v, that is

    \[ \frac{d}{dt}\begin{bmatrix}e_d\\ \Delta v\end{bmatrix} = \begin{bmatrix}-h k_d & 1 - h k_v\\ -k_d & -k_v\end{bmatrix}\begin{bmatrix}e_d\\ \Delta v\end{bmatrix}. \]

The trace is -(k_v + h k_d), the determinant h k_d k_v + k_d(1 - h k_v) = k_d: the characteristic polynomial is

    \[ s^2 + (k_v + h k_d)\,s + k_d = 0, \]

stable for every k_d \mathchar"313E  0, k_v + h k_d \mathchar"313E  0 (the Routh criterion for second order). With h = 1.8 s: natural frequency \sqrt{0.15} = 0.39 rad/s and damping (0.6 + 0.27)/(2\cdot 0.39) = 1.12, that is no overshoot (the second-order response). A slow, calm following is what is wanted: nobody wants a car that reacts to every small change of the vehicle ahead.

Following a vehicle that changes speed
Following a car that oscillates between 85 and 95 km/h with a period of 12 s: our speed (blue) follows with a slight delay and a smaller amplitude; the distance stays close to the desired one.

String stability

A single stable car is not enough. In a string of cars with ACC each one follows the previous one: if each amplifies the speed variations a little, the last one suffers a huge disturbance (it is the wave that appears out of nowhere in traffic jams). With Laplace transforms (see transfer functions), for car i following car i-1:

    \[ s V_i = k_d\left(\frac{V_{i-1} - V_i}{s} - h V_i\right) + k_v (V_{i-1} - V_i) \;\Rightarrow\; G(s) = \frac{V_i}{V_{i-1}} = \frac{k_v s + k_d}{s^2 + (k_v + k_d h)s + k_d}. \]

The string does not amplify disturbances if |G(j\omega)| \le 1 at every frequency. With s = j\omega:

    \[ k_d^2 + k_v^2\omega^2 \le (k_d - \omega^2)^2 + (k_v + k_d h)^2\omega^2 \iff \omega^2 + k_d^2 h^2 + 2 k_v k_d h - 2 k_d \ge 0. \]

The worst case is \omega \to 0; dividing by k_d gives the condition

    \[ \boxed{\,k_d h^2 + 2 k_v h \ge 2\,}. \]

With k_d = 0.15, k_v = 0.6: for h = 1.5 it is 2.14 and fine; for h = 1.0 it is 1.35 and the string amplifies. This is why the gains cannot be fixed: in the code k_v grows when the time gap gets shorter, k_v = \max(0.6,\ (2 - k_d h^2)/(2h) + 0.05), that is 0.98 for h = 1.0 s.

Magnitude of the string transfer function
|G(j\omega)| in decibels for different time gaps with fixed gains (solid lines): for 0.8 and 1.0 s it exceeds 0 dB, the string amplifies. The dashed lines are the same time gaps with k_v adapted: they stay below 0 dB.
A string of eight cars
Eight cars in a string; the first brakes at −2 m/s² for 2 s and then accelerates again. With h = 0.6 s and fixed gains (left) the wave grows: the first drops to 75.6 km/h, the last to 68.9. With h = 1.5 s (right) it is damped: the last drops only to 79.7 km/h. The actuators have a delay of 0.2 s, which worsens the condition: it is one of the reasons for the 0.8 s minimum of the standard.

Approaching a slower vehicle

We arrive at 130 km/h and ahead, at 200 m, there is a car at 80 km/h. The following controller, left alone, waits until the distance error becomes small and then brakes hard: with \Delta v = -14 m/s the term k_v\Delta v is -8.3 m/s², limited to -3.5. An experienced driver does something else: stops accelerating, then brakes gently and steadily.

The kinematic deceleration. In relative motion, with the target at constant speed, a constant deceleration a cancels the speed difference \Delta v after a relative distance D given by \Delta v^2 = 2|a|D. If we want \Delta v to vanish exactly when the distance reaches d_{end}, we need

    \[ a_{kin} = -\frac{\Delta v^2}{2\,(d - d_{end})} + a_t. \]

Recomputed every cycle, this law corrects its own errors: if the car brakes less than planned, the remaining distance shrinks and the request grows.

Elasticity. It is not necessary to arrive exactly at d_{des} with \Delta v = 0. If we accept going temporarily below the desired distance by a margin \delta and recovering it calmly afterwards, the distance available for braking grows and the braking gets softer: with d_{end} = d_{des} - \delta,

    \[ |a_{kin}| = \frac{\Delta v^2}{2\,(d - d_{des} + \delta)} \mathchar"313C  \frac{\Delta v^2}{2\,(d - d_{des})}. \]

In the code \delta = \min(0.25\,h\,v,\ 12\ \mathrm{m}): a quarter of the desired distance, never more than 12 m.

Phases of the approach
The phases in the plane of distance and closing speed: (1) stop accelerating and coast; (2) constant braking on the parabola \Delta v^2 = 2|a|(d - d_{end}), which ends at d_{des} - \delta instead of d_{des} (dashed); (3) a short elastic undershoot; (4) gentle recovery to d_{des}.

Three details make the manoeuvre robust:

  • mode with hysteresis: the approach starts with a real closing speed (above 1 m/s) and ends when it is almost zero or the elastic margin is used up; without hysteresis, with \Delta v close to zero the mode switches on and off;
  • mild phase: as long as a_{kin} is weaker than what releasing the pedal gives (-0.35 m/s²), the following controller does the job, but without braking more than necessary (\max(a_f, a_{kin})) and without closing faster than \sqrt{2\cdot 0.35\,(d - d_{end})};
  • elastic recovery: after the approach the missing distance is recovered with a reduced gain (k_d = 0.08) and a deceleration of at most -0.3 m/s²: the distance opens up because we slow down slightly, not because we brake.
Approach: controller, kinematics, kinematics with elasticity
From 130 to 80 km/h starting at 180 m. The following controller alone (grey) waits 5 s and then brakes down to −1.9 m/s². The kinematic law (blue) brakes right away and steadily, at most −0.94 m/s². With elasticity (green) the braking is even softer, −0.80 m/s²; the distance drops to 29 m against the desired 37 and is then recovered slowly.

Overtaking preboost

We follow a truck at 80 km/h with the set speed at 130. The driver signals left to overtake. A simple ACC keeps the 37 m of distance until the truck has left the corridor, and only then accelerates: the overtake starts slowly, and whoever comes from behind in the overtaking lane has to brake.

Preboost anticipates the overtake: with the indicator towards the overtaking lane, a target slower than the set speed and the next lane free (corner radars and fusion say so), the planner

  • reduces the time gap towards h_{pb} = 0.9 s, always with the slow change of 0.25 s per second;
  • adds an acceleration a_{pb} = 0.6 m/s² to the following term;
  • stops after 4 s if the lane change does not happen, or as soon as the indicator is switched off, and the distance goes back gently to its normal value.
Overtaking preboost
With the left indicator the time gap goes down towards h_{pb} and the car starts accelerating before the truck has left the corridor; after the lane change the target is free and the acceleration rises towards a_{max}.
Overtaking with and without preboost
Indicator at 5 s, lane change at 8 s (violet area). With preboost (violet) at the lane change we are already at 86 km/h instead of 80 and we reach 120 km/h 1.1 s earlier; the distance to the truck goes down to a minimum of 34.6 m (0.9 s of time gap at that speed), never below.

Stop & go

In a queue the target brakes until it stops. The following controller, with the target acceleration fed forward, would copy its braking and stop us where we are, perhaps at 15 m. What we need is to stop d_0 from the point where the other one will stop.

The stopping law. If the target brakes with a_t \mathchar"313C  0, it will stop after the distance v_t^2 / (2|a_t|), that is at the distance x_{stop} = d + v_t^2/(2|a_t|) from us. To stop d_0 behind that point with constant deceleration:

    \[ a_{stop} = -\frac{v^2}{2\,(x_{stop} - d_0)}. \]

If the target is already stopped, x_{stop} = d. The law replaces following and approach when the target will stop within 4 s and we are below 54 km/h. At low speed the wider limits of ISO 22179 apply, interpolated as seen in lesson 1: down to −5 m/s² and up to +4 m/s² below 18 km/h. When we are stopped less than d_0 + 2 m from the stopped target, the planner requests the stop and the state machine goes to STANDSTILL: the brakes hold the car (lesson 6).

Stop and go uphill
Stop & go on a 4% climb with the whole longitudinal chain. The car ahead stops twice. First stop: we stop at 3.93 m and drive off by ourselves about 0.6 s after the other car has moved off. Second stop: after 3 s the state machine goes to WAIT; we drive off when the driver presses RES (33 s). No rollback on the climb: the brakes (red) release only while the engine torque (green) builds up.

Arbitration, limits and jerk

The proposals of the modes (set speed a_{cc} = 0.4\,(v_{set} - v), following, approach, stop) go into the minimum block: the most cautious wins. Then:

  • saturation to the limits of the standard, which depend on speed (from -3.5/+2 m/s² above 72 km/h to -5/+4 below 18 km/h);
  • jerk limit: the request changes by at most 2.5 m/s³ (comfort); in firm braking (below -2 m/s²) up to 5 m/s³ at low speed.
self.mode = min(cands, key=cands.get)                  # the most cautious proposal wins
a_lo, a_hi, j_lo = iso_limits(v, p)                    # ISO 15622 / 22179, speed dependent
a_goal = float(np.clip(cands[self.mode], a_lo, a_hi))
if self.stop_request:
    a_goal, self.mode = min(a_goal, -0.5), "stop"
jerk = j_lo if a_goal < -2.0 else p.jerk_max
self.a_out += float(np.clip(a_goal - self.a_out, -jerk * dt, p.jerk_max * dt))

The alternative: model predictive control

A different approach is model predictive control (MPC), which we saw in the lesson on optimal and predictive control. With the state (e_d, \Delta v, a) and the model of relative motion, every cycle a quadratic problem is solved over a horizon of a few seconds: minimise a weighted sum of distance error, speed difference, acceleration and jerk, with explicit constraints (limits of the standard, minimum distance, set speed). Gentle approach, elasticity and early braking come out of the cost and the constraints by themselves.

Rule-based (this lesson)Predictive (MPC)
behaviourdesigned mode by modecomes from cost and constraints
constraintswith saturations and special casesexplicit, met by construction
computationmicrosecondsa quadratic problem every 20 ms
calibration and certificationevery parameter has a physical meaningless intuitive weights, solver to qualify

Many series systems stay rule-based precisely because of the last row; MPC is used more and more where the ECU has computing power and long predictions are needed (curves, speed limits, lane changes).

The request a_{req} is ready. But the car has no acceleration command: it has an engine and brakes. In the next lesson we turn it into wheel torque, with all the physics of the vehicle.

Subscribe
Notify of
guest
0 Comments
Oldest
Newest Most Voted
0
Would love your thoughts, please comment.x
()
x