MPC Path Tracking Lab
Two QP-based model predictive controllers steer identical vehicles through identical steering actuators. Case A carries a second-order steering model inside its prediction; Case B assumes the wheels follow the command instantly. Vary speed, horizons, input weight and actuator bandwidth to compare path error, steering oscillation and constraint activity, then check the trend with a speed sweep.
| Metric | A · aware | B · vehicle-only | A vs B |
|---|---|---|---|
| RMS lateral error ey | — | — | — |
| Peak lateral error |ey| | — | — | — |
| RMS heading error eψ | — | — | — |
| Mean steering rate |δ̇| | — | — | — |
| Peak steering rate max |δ̇| | — | — | — |
| Peak lateral accel. ay | — | — | — |
| Constrained QP steps share | — | — | — |
| Hildreth iterations mean | — | — | — |
Full-course batch runs with the current Np, Nc, r and actuator. Hollow markers: the vehicle left the path (peak |ey| > 1 m).
A x = [ey, ėy, eψ, ėψ, δ, δ̇]T
δ̈ = ωn2(δcmd − δ) − 2ζωnδ̇
B x = [ey, ėy, eψ, ėψ]T, δ ≡ δcmd
2-DOF bicycle error dynamics, curvature preview κ(s) as a measured disturbance, ZOH at Ts.
minΔU Σk=1Np (qyey,k2 + qψeψ,k2) + r Σj=0Nc−1 Δuj2
s.t. |uj| ≤ δmax, |Δuj| ≤ δ̇maxTs
A only: |δk| ≤ δmax, |δ̇k| ≤ δ̇max
Condensed to a dense QP in Nc increments; receding horizon applies Δu0.
Hildreth's procedure: coordinate ascent on the dual, entered only when the unconstrained optimum violates a bound.
Plant: nonlinear single-track model, m = 1530 kg, tanh tyre saturation (μ = 1.0), actuator angle/rate saturation. Neither controller's model matches it exactly.
Avg. solve — ms per controller step (this browser).
What the paper validated
The paper integrated an identified second-order steering model with a 2-DOF vehicle model, then formulated QP-based MPC with steering angle and rate constraints. DLC and road-course simulations used Np=20, Nc=8, Ts=10 ms, r=60 as the baseline configuration. The vehicle-only predictor developed increasing oscillation as speed rose.
What this browser model reproduces
The controllers follow the paper's structure (2-DOF error model plus second-order steering model, angle and rate constraints, Hildreth QP) and solve a real QP every step. The plant is a nonlinear single-track model with tyre saturation rather than 27-DOF CarSim, and the browser runs at Ts = 20 ms. Read the numbers as trends, not as a reproduction of the paper's figures.
View the publication on About →