Technical Summary: Adaptive Differentiable Model Predictive Contouring Control (AD-MPCC) for Autonomous Racing

1. Problem Statement

Autonomous racing requires vehicles to operate at their physical limits while optimizing performance objectives such as lap-time minimization. While Model Predictive Contouring Control (MPCC) has become a standard framework for balancing tracking accuracy and racing speed, it faces significant challenges in dynamic environments, particularly on tracks with varying road-surface conditions.

The primary difficulties addressed in this work are:

1. Model Uncertainty: The dominant source of uncertainty in racing arises from tire-road interactions, which vary with surface conditions (e.g., asphalt vs. gravel). Standard physics-based models often assume fixed parameters, leading to inaccuracies when surface friction changes.
2. Objective Weight Adaptation: In multi-surface scenarios, updates to the dynamic model (e.g., tire friction parameters) necessitate corresponding adjustments to the MPCC objective weights to maintain safety and performance. However, manually tuning these weights is impractical, and standard differentiable MPC techniques do not directly apply to MPCC due to the implicit coupling between the reference trajectory and the optimized progress variable.
3. Real-Time Constraints: While differentiable programming offers a mechanism for weight optimization, the computational burden of solving high-dimensional optimization problems online is often prohibitive for real-time racing applications.

2. Methodology: The AD-MPCC Framework

The authors propose the Adaptive Differentiable Model Predictive Contouring Control (AD-MPCC) framework, which integrates two core components to address the above challenges: online parameter estimation and differentiable weight adaptation.

A. Online Pacejka Parameter Estimation

To handle varying road surfaces, the framework employs a Regularized Moving Horizon Estimation (MHE) scheme.

• Model: The vehicle dynamics are modeled using a dynamic bicycle model where lateral forces are governed by the Pacejka "Magic Formula."
• Estimation: The system estimates the Pacejka parameters ($\eta$, including friction coefficients $B, C, D$) in real-time using a dataset of recent vehicle states and inputs.
• Mechanism: The MHE objective function includes:
  • Exponentially decaying weights: This prioritizes recent observations to rapidly capture transitions between different road surfaces.
  • Prior regularization: The estimation is regularized toward nominal parameters ($\eta_{gt}$) derived from standard road conditions to ensure stability.
• Output: The estimated parameters $\eta_k$ are updated at every time step and fed into the predictive model.

B. Differentiable MPCC (Diff-MPCC)

To automatically tune the MPCC objective weights ($q = [q_c, q_l, q_v]^\top$) based on the estimated dynamics, the authors develop a Diff-MPCC approach.

• Concept: Instead of fixed weights, the framework optimizes the weights to minimize a long-horizon performance cost $\ell$.
• Differentiation: The method computes gradients of the performance cost with respect to the weights ($\partial \ell / \partial q$) by differentiating through the MPCC optimization problem.
• Technical Challenge: Unlike standard MPC, MPCC involves an implicit progress variable $\theta$ that couples the reference trajectory to the optimization. This requires higher-order derivatives of the reference path.
• Solution: The authors apply the Implicit Function Theorem (IFT) to the Karush-Kuhn-Tucker (KKT) optimality conditions of the MPCC problem. By approximating the solution via a barrier method and solving the resulting linear system, they compute the necessary gradients to update the weights via a projected gradient descent method.

C. Pacejka-Informed Machine Learning (PaIML)

Since running the full Diff-MPCC optimization loop online is computationally expensive, the authors propose a PaIML technique for real-time deployment.

• Offline Training: A supervised learning model is trained on data generated by the Diff-MPCC solver. The dataset consists of vehicle states and estimated parameters paired with the optimal weights computed by Diff-MPCC.
• Physics-Informed Reduction: To reduce the input dimensionality from 14 to 5, the authors leverage the physical structure of the Pacejka model. Instead of feeding raw parameters, the model uses:
  1. Friction coefficients ($\mu_r, \mu_f$) derived from the Pacejka parameters.
  2. Longitudinal and lateral velocities ($v_x, v_y$).
  3. The progress variable ($\theta$).
• Architecture: An XGBoost model is used to map these 5 inputs to the optimal MPCC weights, enabling fast, near-optimal inference during real-time operation.

3. Key Contributions

The paper highlights three primary contributions:

1. AD-MPCC Framework: A novel control architecture that jointly performs online estimation of the vehicle dynamics model and automatic adaptation of MPCC objective weights for autonomous racing on varying surfaces.
2. Prior-Regularized MHE: A specific estimation scheme that utilizes exponentially decaying weights to rapidly estimate Pacejka parameters, ensuring the predictive model adapts quickly to surface transitions.
3. Differentiable MPCC & PaIML:
   • Development of a differentiable programming technique for MPCC that enables gradient-based weight tuning.
   • Design of a real-time approximation method (PaIML) that reduces computational complexity by leveraging physical insights to lower input dimensionality.

4. Simulation Results

The framework was validated in the F1TENTH-Gym environment using an Oschersleben racetrack under two scenarios:

A. Single-Surface Scenario (Known $\mu_{max} = 1.2$)

• Performance: Both Diff-MPCC and AD-MPCC significantly reduced lap times compared to the baseline MPCC.
  • Diff-MPCC: Achieved the fastest lap time (64.08 s vs. 75.57 s for baseline).
  • AD-MPCC: Achieved a lap time of 64.89 s with a smaller mean lateral offset (0.926 m) compared to Diff-MPCC (1.040 m), indicating a better trade-off between speed and safety.
• Computation: All methods operated within the 50 ms sampling budget, with AD-MPCC requiring approximately 24.18 ms per step.

B. Multi-Surface Scenario (Unknown $\mu_{max}$ varying from 0.7 to 1.2)

• Safety & Robustness: This scenario proved critical. The baseline MPCC, A-MPCC (estimation only), and Diff-MPCC (weight adaptation only) all crashed at the first slippery segment ($\mu_{max} = 0.75$).
• AD-MPCC Success: AD-MPCC was the only controller capable of completing the lap safely.
• Mechanism: The online estimation successfully detected the drop in friction (e.g., parameter $D_r$ decreased from 1.55 to ~1.45), and the weight adaptation adjusted the controller to maintain stability, achieving an average velocity of 11.65 m/s with a lateral offset of 1.069 m.

5. Significance and Claims

The paper claims that AD-MPCC represents a significant advancement in autonomous racing control by addressing the dual challenges of model uncertainty and controller adaptability.

• Safety in Uncertainty: The framework is the only method tested that ensures safe vehicle operation in unknown, multi-surface environments where other controllers fail.
• Performance: On known tracks, the framework achieves lap times approximately 11 seconds faster than the standard MPCC.
• Practicality: By combining differentiable optimization with physics-informed machine learning, the authors demonstrate that complex, adaptive control strategies can be deployed in real-time without exceeding computational budgets.

The authors note that the current validation relies on simulator-generated data (F1TENTH-Gym) and that future work will focus on real-world deployment and deeper theoretical analysis of stability and robustness.