High-Speed Forward and Reverse Tractor-Trailer Driving with Learned Vehicle Dynamics

Aaron Chen, David Shen, Taekyung Kim, Kaleb Ben Naveed, Dimitra PanagouarXiv 2026 · arXiv

TL;DR

Researchers have developed a predictive control framework that allows tractor-trailers to navigate complex tracks at high speeds in both forward and reverse. By swapping out traditional, rigid physics equations for a neural network that "remembers" recent vehicle behavior, the system achieves significantly better stability and speed tracking than standard analytical models, even on unseen snowy or sloped terrain.

What Problem Does This Paper Address?

Driving a tractor-trailer is a nightmare for control engineers. These systems are highly nonlinear and inherently unstable, particularly in reverse—think of the "jackknife" effect where the trailer swings out of control. At high speeds, the problem intensifies as tire slip, road friction, and suspension dynamics become critical.

Traditional Model Predictive Control (MPC) relies on an analytic physics model to "look ahead" and plan moves. However, these models often struggle because they can't easily account for hidden variables like changing road friction or the exact way a suspension compresses under load. If the model’s prediction is wrong, the controller’s plan fails, leading to stalls or crashes.

Methods

The authors propose a framework combining Model Predictive Path Integral (MPPI) control with a learned dynamics model.

  1. The Controller (MPPI): Unlike gradient-based MPC that solves complex optimization problems, MPPI is a sampling-based method. It "imagines" thousands of random control sequences, simulates them using the dynamics model, and calculates a weighted average of the best-performing ones to decide the next move.
  2. The Learned Model: Instead of using the standard Fiala tire force equations (the analytic baseline), the team trained a Multilayer Perceptron (MLP). Crucially, this MLP is history-conditioned: it takes a window of the last four state-action pairs as input. This allows the network to implicitly "sense" environmental factors like low friction or slopes by observing how the vehicle responded to recent commands.
  3. The Analytic Baseline: For comparison, the authors derived a reduced-order tractor-trailer model using a bicycle model for the tractor and a single-axle model for the trailer, incorporating Fiala brush tire forces to handle high-speed slip.

Datasets

The researchers used the BeamNG.tech simulator, a high-fidelity physics engine popular in the automotive industry.

  • Training Data: They collected trajectories on a flat asphalt track with varying friction levels ($\mu$ from 0.2 to 1.0) and speeds ranging from -50 km/h (reverse) to 90 km/h (forward).
  • Evaluation Environments: They tested the controllers on four distinct scenarios based on the Circuit de Barcelona-Catalunya layout:
  • Asphalt: Flat, high friction.
  • Snow: Irregular low-friction patches.
  • Slope: Elevation changes and banking (not present in training data).
  • Snow + Slope: The ultimate stress test combining both.

Key Insights and Ideas

  • History is a Proxy for Hidden States: The most significant "aha" moment is that a simple history window ($H=4$) allows the neural network to outperform an analytic model even when that analytic model is given the ground-truth friction coefficient. The learned model captures subtle, unmodeled effects (like suspension roll or powertrain lag) that physics equations often simplify away.
  • Generalization to Unseen Terrain: Even though the model was never trained on hills or banked turns, it successfully transferred to the "Slope" environment. This suggests the network learned a robust relationship between control inputs and vehicle response that transcends flat-ground assumptions.
  • Reverse Stability: The learned model was particularly effective at regulating the "hitch angle." By accurately predicting how the trailer would swing, the MPPI controller could make preemptive steering corrections to prevent jackknifing at speeds where human drivers would likely fail.

Key Results

The learned MPPI (Controller 1) was compared against an analytic MPPI, LQR-PID, and two other MPC variants.

  • Completion Rates: In reverse configurations across all environments, the learned model completed 62 out of 72 runs. The analytic baseline only managed 41 out of 72, frequently stalling or jackknifing.
  • Speed Tracking: The learned model maintained mean speeds much closer to the targets. For a -50 km/h reverse setpoint on asphalt, it achieved 47.2 km/h, while the analytic model lagged at 22.7 km/h.
  • Hitch Regulation: The learned model consistently showed lower Root-Mean-Square (RMS) hitch angles, indicating a much "smoother" and more stable trailer motion.
  • Prediction Accuracy: In a 5-second "open-loop" prediction test, the learned model's position error was often 2-4 times lower than the analytic models.

Limitations and Future Work

While impressive, the system isn't perfect. In the "Snow + Slope" environment, the controller still suffered from stalls (where the vehicle stops making progress) in about half of the reverse runs. The authors also noted that the model was trained and tested with a consistent trailer; changing the trailer's payload or length would likely require retraining or an explicit adaptation module.

Future work aims to explore these sources of uncertainty, such as varying cargo weights, to make the system truly "plug-and-play" for any fleet.

Summary updated

Abstract

Tractor-trailer systems are difficult to control, especially unstable during reverse motion, and their dynamics depend on difficult-to-measure factors, including road friction and suspension coefficients. We introduce a learned dynamics model for tractor-trailer control and compare it with a reduced-order dynamics model with Fiala tire forces as an analytic baseline. Specifically, we compare a model predictive path integral (MPPI) controller that uses the learned dynamics model to an MPPI controller that uses the analytic model, as well as to three additional baseline controllers. In the main closed-loop tracking experiments across four simulated environments, MPPI with learned dynamics achieves mean speeds closer to the target magnitudes and lower root-mean-square (RMS) hitch angles than MPPI with the analytic model at every tested setpoint, with higher completion rates in seven of eight reverse configurations.
BibTeX
@misc{chen2026high,
  author = {Aaron Chen and David Shen and Taekyung Kim and Kaleb Ben Naveed and Dimitra Panagou},
  title = {{High-Speed Forward and Reverse Tractor-Trailer Driving with Learned Vehicle Dynamics}},
  year = {2026},
  eprint = {2611.20324},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20324},
}
Figure