Data-Driven Predictive Control for Autonomous Vehicle Racing under Varying Friction

Julius Beerwerth, Julius Tischbein, Bassam AlrifaeearXiv 2026 · arXiv

In the high-stakes world of autonomous vehicle racing, tire-road friction is the ultimate "invisible variable." A sudden patch of water or a change in tire temperature can turn a perfect racing line into a trip to the barrier. Traditionally, engineers handle this by building complex online estimators to guess the friction coefficient ($\mu$) and adapt their control models accordingly.

A recent paper, "Data-Driven Predictive Control for Autonomous Vehicle Racing under Varying Friction," suggests a simpler, more robust path: just pool all your data together and let the optimization math do the heavy lifting.

TL;DR

Instead of explicitly estimating tire friction, this paper shows that Data-Driven Predictive Control (DPC) can achieve near-optimal performance by "pooling" trajectory data from multiple friction levels. The controller implicitly favors data that matches the vehicle's recent history, providing a form of "partial adaptation" that rivals complex nonlinear model-based controllers.

What Problem Does This Paper Address?

Autonomous racing controllers typically rely on Model Predictive Control (MPC). However, MPC is only as good as the underlying vehicle model. Tire parameters change drastically with friction; a model identified on a dry track will fail on a wet one.

Data-Driven Predictive Control (DPC) offers an alternative by predicting future motion directly from recorded input-output trajectories. But DPC has a similar "matching" problem: if your database only contains dry-track data, it won't accurately predict behavior on a slippery surface. The standard solution—identifying the current friction and switching to a "matched" dataset—requires a dedicated, often brittle, adaptation layer. The authors ask: Can we just throw all the data (dry and wet) into one big pool and still get accurate results?

Methods

The researchers utilize a regularized DPC formulation. At its core, DPC predicts future outputs by finding a linear combination of previously recorded trajectories that matches the vehicle's recent "initial trajectory" (the past few time steps).

The "secret sauce" explored here is Data Pooling. They compare three variants:

  1. Matched DPC (The Oracle): Uses only data recorded at the current friction level.
  2. Mismatched DPC: Uses data from a different friction level (e.g., using "dry" data on a "wet" track).
  3. Pooled DPC (The Proposed): Combines data from both friction levels into one dataset.

To keep the computation efficient, they use Contextual Sampling to select a relevant subset of trajectories online based on the current state and reference path. Crucially, they provide a theoretical analysis showing that the DPC's regularization term naturally penalizes trajectories that don't align with the recent history, effectively "filtering" the pool for the most relevant friction data without needing an explicit estimator.

Datasets

The approach was validated in two distinct environments:

  • Simulation: A high-fidelity, full-scale multibody vehicle model. Data was collected at $\mu=0.5$ (low friction) and $\mu=1.1$ (high friction), totaling 2000 seconds of driving near the physical limits of grip.
  • Hardware (CPM Lab): 1:18 scale vehicles. High friction was achieved with rubber tires, while low friction was simulated by covering the same tires with a smooth film.

Key Insights and Ideas

  • The "Informative Past": The vehicle's recent history (the initial trajectory) acts as a signature of the current friction. If the car is sliding more than expected for a given steering input, the DPC optimization will naturally assign higher weights to "low-friction" trajectories in the pool because they provide a better fit for that recent behavior.
  • Implicit Adaptation: Pooled DPC provides "partial adaptation." It isn't perfect—there's a slight lag at maneuver entries or friction transitions—but it is far more robust than using mismatched data.
  • Regularization vs. Selection: Interestingly, the researchers found that contextual sampling doesn't perfectly separate the data by friction. Instead, the regularization within the optimization problem does the heavy lifting, shifting the prediction toward the matched subset.
Distribution of prediction errors showing Pooled DPC (top) performing much closer to Matched DPC (middle) than Mismatched DPC (bottom).

Key Results

  • Prediction Accuracy: In simulations, Pooled DPC kept the multi-step root-mean-square (RMS) prediction error within 1.5 times that of the "Oracle" Matched DPC. In contrast, Mismatched DPC was 2.2 to 3.4 times worse.
  • Competitive Tracking: Pooled DPC achieved a median tracking error of 13.7 cm, nearly identical to a state-of-the-art Nonlinear MPC (NMPC) baseline of 13.3 cm.
  • Hardware Success: In real-world lab experiments, Pooled DPC performed remarkably close to Matched DPC, with tracking errors differing by only 0.1 cm to 0.2 cm across different tire sets.
Comparison of tracking and prediction errors along the racetrack. Pooled DPC (green) stays much closer to the zero-line (Matched DPC) than Mismatched DPC (purple).

The researchers also visualized the "implied friction"—the friction level the controller thinks it's seeing based on the data it uses. As shown in the figure below, the controller correctly identifies high and low friction zones in corners where the vehicle dynamics are most influenced by grip.

Visualization of true track friction (inner ring) vs. the friction implied by the Pooled DPC prediction (outer ring).

Limitations and Future Work

While Pooled DPC is powerful, it isn't magic.

  • Data Coverage: The method relies on having data that covers the demanded speeds and curvatures. If you try to drive faster than any trajectory in your pool, the system fails.
  • Transition Lag: There is a brief period after a friction change where the "initial trajectory" hasn't yet captured the new dynamics, leading to temporary prediction errors.
  • Future Directions: The authors suggest investigating "subset-aware regularization" to more explicitly handle pooled data and improving trajectory selection methods to better distinguish between operating conditions.

This work demonstrates that for complex, nonlinear systems like racing cars, sometimes the best way to handle uncertainty isn't to model it, but to let the data speak for itself.

Summary updated

Abstract

Data-driven predictive control (DPC) predicts future outputs by combining recorded input-output trajectory data such that the combined trajectory approximately matches a measured initial trajectory and tracks a provided reference. This paper investigates to what extent and under which conditions DPC using data pooled across operating conditions retains the prediction accuracy of DPC using data from the current operating condition. In autonomous vehicle racing, the tire-road friction varies between racetrack sections and tire sets. If the initial trajectory and the recorded data originate from different frictions, prediction accuracy of DPC degrades. Ensuring the frictions match requires knowing the friction during recording and estimating the current friction explicitly. Instead, we pool data recorded at different frictions. We compare the prediction accuracy of DPC using data recorded under the current friction (matched), a different friction (mismatched) and data pooled from both frictions. We evaluate in simulation and small-scale vehicle experiments in the Cyber-Physical Mobility Lab (CPM Lab). Our results show that pooled data keep the multi-step root-mean-square prediction error within 1.5 times that of matched data, whereas mismatched data increase it by a factor of 2.2 to 3.4. To assess how this prediction accuracy carries over to control, we evaluate the tracking error from the raceline in closed loop. In simulation, DPC with pooled data tracks comparably to nonlinear model predictive control (13.7 cm versus 13.3 cm) and close to DPC with matched data in the CPM Lab, differing by 0.1 cm on the low-friction tires (1.6 cm versus 1.5 cm) and by 0.2 cm on the high-friction tires (0.9 cm versus 1.1 cm). DPC thus provides partial adaptation to varying friction, without a dedicated adaptation method.
BibTeX
@misc{beerwerth2026data,
  author = {Julius Beerwerth and Julius Tischbein and Bassam Alrifaee},
  title = {{Data-Driven Predictive Control for Autonomous Vehicle Racing under Varying Friction}},
  year = {2026},
  eprint = {2611.21378},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21378},
}
Figure