⚡ electrical engineering

CAPE: Control Algorithm Performance Evaluation under Learned Vehicle Dynamics Models

The paper introduces the CAPE framework for systematically benchmarking racing controllers under learned vehicle dynamics models, demonstrating that the proposed Enhanced Physics Model (EPM) significantly outperforms state-of-the-art alternatives in lap time, tracking accuracy, and robustness against realistic uncertainties.

Original authors: Malik Ali, Musabbir Ahmed Arrafi, Nicholas M. Stiffler, Krishna Bhavithavya Kidambi

Published 2026-06-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Malik Ali, Musabbir Ahmed Arrafi, Nicholas M. Stiffler, Krishna Bhavithavya Kidambi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are coaching a team of race car drivers. Some drivers are old-school veterans who rely on a simple, hand-drawn map of the track. Others are tech-savvy rookies using a high-tech GPS that learns from data. Your goal is to figure out which map helps the drivers go the fastest and safest, especially when the track gets slippery, the wind blows, or the car's engine acts up.

This paper introduces a new "coaching system" called CAPE. It's a testing ground designed to see how different "maps" (vehicle models) affect different "drivers" (control algorithms) when they are racing a self-driving car.

Here is the breakdown of the paper's story:

1. The Problem: The Map vs. The Reality

In autonomous racing, the car needs a mathematical "map" of how it moves.

Old Maps (Classical Models): These are like simple sketches. They work okay on a straight, dry road, but if the car starts sliding or the tires get hot, the sketch becomes wrong. The car might crash because the map didn't predict the slide.
New Maps (Learned Models): These are like GPS systems that learn from real driving data. They are smarter, but researchers weren't sure which specific "smart GPS" worked best when paired with different driving styles.

2. The Three "Smart GPS" Models Tested

The authors compared three different ways the car learns its own physics:

DPM (Deep Pacejka Model): Think of this as a GPS that knows the rules of tire friction very well but is a bit rigid. It's like a driver who knows the theory perfectly but can't adapt quickly to a sudden puddle.
DDM (Deep Dynamics Model): This is a GPS that uses a massive neural network (a giant brain) to guess how the car moves. It's very flexible but can be "heavy" and slow to compute, like a driver who overthinks every turn.
EPM (Enhanced Physics Model - The Winner): This is the authors' new invention. Imagine a GPS that combines a solid understanding of physics (like how weight shifts when you turn) with a smart, lightweight brain that learns the specific quirks of the car. It's the "Goldilocks" model—not too simple, not too complicated, but just right.

3. The Five "Drivers" (Controllers)

To make the test fair, they didn't just test the maps; they tested them against five different types of drivers:

The Classic Driver (PID): Follows the rules strictly. Good for smooth roads, bad for chaos.
The Tough Driver (RISE): Built to handle bumps and wind.
The Planner (Nominal NMPC): Plans the whole lap ahead based on the map.
The Adaptive Driver (Adaptive NMPC): Changes its mind on the fly if the car feels different.
The Safety-First Driver (Robust NMPC): Assumes the worst and plans for errors.

4. The Race: Adding Chaos

The real test wasn't just driving on a perfect track. The authors added "chaos" to the simulation to see who could handle it:

Noise: Like trying to drive with foggy windows (sensor errors).
Delays: Like the steering wheel taking a split second to turn (actuator lag).
Wind: A crosswind pushing the car sideways.
Weight Changes: The car getting heavier or the tires getting softer.

They turned up the "chaos dial" (called $\eta$ ) to see at what point the drivers would crash or fail to finish the lap.

5. The Results: Who Won?

The results were clear: The Enhanced Physics Model (EPM) won.

Speed: When paired with the best driver (Adaptive NMPC), the EPM car finished a lap in 5.82 seconds. The other models took much longer (8.80s and 12.99s) or crashed entirely.
Accuracy: The EPM car stayed much closer to the ideal racing line. It made fewer mistakes in steering and speed.
Resilience: As the chaos increased (more wind, more noise), the other models started to fail. The EPM car kept going, completing laps even when the others gave up.

The Big Takeaway

The paper proves that you can't just look at a controller (the driver) in isolation. A great driver with a bad map will crash. A great map with a bad driver won't go fast.

The CAPE framework showed that by using their new Enhanced Physics Model (EPM), you get a car that is not only faster but also much tougher. It handles the "bad weather" of real-world racing (noise, delays, and uncertainty) better than the previous state-of-the-art models.

In short: If you want a self-driving race car that can handle the unexpected without crashing, you need a model that understands both the physics of the car and the messy reality of the track. That's exactly what EPM does.

Technical Summary: CAPE – Control Algorithm Performance Evaluation under Learned Vehicle Dynamics Models

1. Problem Statement

Autonomous racing operates near the physical limits of vehicle dynamics, where rapid acceleration, aggressive cornering, and nonlinear tire behavior render simplified classical models (e.g., standard bicycle models) insufficient. Small modeling errors in these regimes can destabilize closed-loop control performance. While recent research has introduced data-driven and hybrid learned dynamics models (such as the Deep Pacejka Model [DPM] and Deep-learning Dynamics Model [DDM]) to improve predictive accuracy, there is a lack of systematic benchmarks evaluating how the choice of a learned dynamics model impacts the performance of heterogeneous closed-loop controllers. Existing studies often evaluate controllers under fixed models or models under fixed controllers, failing to isolate the specific contribution of the dynamics model to robustness and tracking accuracy across diverse control architectures.

2. Methodology

The authors propose the Control Algorithm Performance Evaluation (CAPE) framework, a systematic methodology designed to benchmark autonomous racing controllers under learned vehicle dynamics models. The framework operates through the following components:

Learned Dynamics Models: The study compares three distinct models:
1. Deep Pacejka Model (DPM): Embeds a Pacejka tire-force layer within a learning architecture while retaining a bicycle-model structure.
2. Deep Dynamics Model (DDM): Uses a physics-constrained neural network to estimate Pacejka coefficients and drivetrain resistances within a single-track model.
3. Enhanced Physics Model (EPM): The proposed model, which combines a dynamic single-track representation with learned parameters for load-sensitive tire forces, longitudinal load transfer, and actuator dynamics. It employs a compact neural parameter estimator (using GRUs and MLPs) to predict physically interpretable coefficients rather than raw states.
Control Architectures: Five heterogeneous controllers are evaluated against each model:
- PID Control
- Robust Inverse of Signum Error (RISE) Control
- Nominal Nonlinear Model Predictive Control (N-NMPC)
- Robust NMPC (R-NMPC)
- Adaptive NMPC (A-NMPC)
Disturbance-Aware Simulation Pipeline: To emulate realistic racing conditions, the framework injects a global disturbance scaling factor ( $\eta$ $η$ ) that simultaneously increases:
- Multiplicative measurement noise (state feedback).
- Process disturbances (unmodeled dynamics).
- Actuator delays (steering latency).
- IMU bias drift.
- Crosswind perturbations.
- Parametric uncertainty (mass, tire stiffness, friction limits).
Experimental Setup: Training and testing are performed on separate racetracks (Track 1 for training, Track 2 for testing) to assess generalization. Performance is measured via longitudinal/lateral tracking errors, control effort, and lap completion times, with "failure" penalized to ensure robustness is a primary metric.

3. Key Contributions

The paper outlines three primary contributions:

Unified Benchmarking Framework: A systematic pipeline (CAPE) that evaluates heterogeneous autonomous racing controllers under three different learned dynamics models (DPM, DDM, and the proposed EPM).
Disturbance-Aware Evaluation: A methodology that systematically injects sensing, actuation, and modeling uncertainties to rigorously study controller robustness, moving beyond idealized simulation conditions.
Empirical Model Comparison: A comprehensive comparison demonstrating how model structure influences tracking accuracy, control effort, and stability. The study isolates the effect of the dynamics model from the controller by keeping controller parameters fixed while switching models.

4. Results

The closed-loop experiments yield the following key findings:

Superiority of EPM: Across all models and controllers, the proposed EPM achieves the best average lap times and the lowest tracking errors. Specifically, the Adaptive NMPC with EPM achieved a lap time of 5.82 s, compared to 12.99 s for DPM and 8.80 s for DDM.
Robustness under Uncertainty: Under a moderate global disturbance scaling factor ( $\eta=1$ $η = 1$ ), EPM significantly outperformed the baselines:
- Reduced longitudinal tracking error by 29.0% (vs. DPM) and 17.2% (vs. DDM).
- Reduced lateral tracking error by 24.6% (vs. DPM) and 12.3% (vs. DDM).
- Increased average velocity magnitude by 39.9% (vs. DPM) and 3.1% (vs. DDM).
Failure Rates: In the uncertainty sweep ( $\eta \in [0, 5]$ ), the EPM paired with NMPC variants (Nominal, Robust, Adaptive) completed all 33 evaluations with zero failures. In contrast, DPM and DDM accumulated 19 and 12 failures, respectively, across the same configurations.
Model Complexity vs. Performance: Despite having more parameters than DPM, EPM demonstrated lower inference latency (0.063 ms vs. 0.081 ms) and higher efficiency. The study notes that while DPM performs well in clean conditions, its performance degrades rapidly with measurement noise, whereas EPM's load-sensitive and actuator-aware parameterization maintains stability.
Ablation Study: An analysis isolating the internal controller model from the propagation model confirmed that the EPM configuration (both as the internal model and the propagation model) yields the tightest raceline adherence and fastest lap times.

5. Significance and Claims

The paper claims that CAPE establishes a necessary systematic benchmark for evaluating learned vehicle dynamics models within a closed-loop control framework. The authors assert that their work demonstrates that improvements in the learned vehicle dynamics representation (specifically incorporating load-sensitive forces and actuator dynamics) translate directly into measurable gains in closed-loop controller robustness and performance.

The study concludes that while classical models diverge from real behavior near handling limits, and while existing learned models improve prediction, the proposed EPM significantly enhances controller robustness under realistic uncertainties. The authors position this work as a step toward validating how learned dynamics affect stability and performance across heterogeneous control architectures, noting that future work will extend the framework to reinforcement learning policies and experimental validation on physical platforms.