Technical Summary: Robust Path Tracking for Vehicles via Continuous-Time Residual Learning: An ICODE-MPPI Approach

1. Problem Statement

Autonomous Ground Vehicles (AGVs) operating in unstructured or disturbed environments require precise path tracking and proactive obstacle avoidance. While Model Predictive Control (MPC) is a standard paradigm, its gradient-based nature often struggles with non-convex cost landscapes. Model Predictive Path Integral (MPPI) control has emerged as a robust, derivative-free, sampling-based alternative capable of handling complex nonlinear dynamics and constraints.

However, the efficacy of MPPI is fundamentally limited by model-plant mismatch. The control law relies on an expectation over sampled trajectories; if the nominal forward dynamics model is inaccurate, the resulting rollouts are biased. This leads to suboptimal control sequences, persistent tracking offsets, and high-frequency control chattering. Existing data-driven solutions, such as Multi-Layer Perceptrons (MLPs) and Gaussian Processes (GPs), face challenges: MLPs often lack physical consistency in discrete-time transitions, while GPs incur prohibitive computational overhead for real-time optimization. Furthermore, integrating continuous-time, stability-guaranteed learning methods with the stochastic sampling architecture of MPPI remains an open challenge.

2. Methodology: The ICODE-MPPI Framework

The authors propose ICODE-MPPI, a hybrid framework that integrates Input Concomitant Neural Ordinary Differential Equations (ICODEs) into the MPPI prediction horizon to learn and compensate for unmodeled residual dynamics.

A. ICODE Architecture for Residual Modeling

Unlike standard Neural ODEs that treat inputs as unknown parameters, the proposed ICODE architecture imposes a control-affine prior on the learned dynamics. This ensures the residual term aligns with the physical structure of robotic systems where control inputs enter state derivatives linearly. The learned residual dynamics, $f_{res}$, are defined as:

$$\dot{x}_{res}(t) = f_{\theta}(x(t)) + \sum_{j=1}^{m} g_{\theta,j}(x(t))u_j(t)$$

Where:

• $f_{\theta}(x)$ represents state-dependent autonomous drift (e.g., unmodeled friction, drag).
• $g_{\theta,j}(x)$ represents control-dependent gain fields (e.g., actuator effectiveness).

This affine decomposition provides a strong inductive bias, improving sample efficiency and generalization compared to fully connected black-box networks. The total predicted dynamics used for MPPI rollouts combine the nominal vehicle model ($f_{nom}$) and the learned residual:
$$\dot{x}_{pred}(t) = f_{nom}(x(t), u(t)) + f_{res}(x(t), u(t); \theta)$$

B. Integration with MPPI

The ICODE model is embedded directly into the MPPI forward prediction module. During the sampling phase, the controller simulates thousands of potential trajectories using the augmented Nominal-ICODE dynamics. This allows the system to anticipate environmental disturbances and unmodeled forces in real-time, proactively compensating for residual errors rather than reacting to them after they occur.

C. Iterative Data Aggregation and Training

To ensure high-fidelity modeling across the operational envelope, the authors employ an iterative data aggregation framework:

1. Random Exploration ($D_{rand}$): The vehicle is excited with random control inputs to capture broad physical responses and stability boundaries.
2. Task-Specific Data ($D_{task}$): As training progresses, the MPPI controller uses the current model to execute path-following tasks, collecting on-policy state-action transitions.
3. Training Loop: The dataset $D = D_{rand} \cup D_{task}$ is used to minimize multi-step prediction error via an RK4 solver. The updated model is redeployed into the MPPI planner, creating a self-correcting feedback loop that continuously reduces modeling uncertainty.

3. Key Contributions

The paper outlines three primary contributions:

1. Residual Dynamics Framework: Development of a continuous-time residual learning framework based on ICODEs and iterative data aggregation, which captures uncertainties while maintaining high physical consistency.
2. Proactive Compensation: Integration of the trained ICODE model into the MPPI prediction horizon, enabling the system to compensate for environmental disturbances and mitigate steady-state tracking drift inherent in nominal models.
3. Comprehensive Benchmarking: A rigorous evaluation of ICODE-MPPI against standard MPPI baselines, utilizing state-variable Root Mean Square Error (RMSE) and control rate probability densities to demonstrate superior tracking precision and control smoothness.

4. Experimental Results

The framework was evaluated in high-fidelity simulations on three complex trajectories (Ellipse, Sine-wave, and Figure-8) under persistent, time-varying external disturbances (modeled as composite sinusoidal signals affecting position and heading).

A. Tracking Precision

• Error Reduction: ICODE-MPPI achieved a significant reduction in cross-tracking errors. In the Ellipse trajectory, the Y-direction error was reduced by up to 69% compared to standard MPPI.
• Drift Suppression: While standard MPPI exhibited persistent steady-state offsets and severe overshooting at curvature inflection points, ICODE-MPPI maintained tight adherence to the reference path by anticipating lateral disturbances.
• Trade-off Analysis: The authors note a slight increase in Yaw (heading) RMSE for ICODE-MPPI in some cases. This is attributed to a deliberate control strategy: the controller prioritizes strict spatial (X, Y) safety against strong lateral disturbances by performing more aggressive heading adjustments, sacrificing perfect heading alignment for positional accuracy.

B. Control Smoothness and Stability

• Chattering Reduction: Analysis of control jerk (rates of change for acceleration and steering) revealed that standard MPPI exhibited wide distributions and flat peaks, indicating severe control chattering. In contrast, ICODE-MPPI demonstrated a higher peak at zero and a narrower distribution for steering rates, signifying significantly smoother steering commands and reduced chattering.
• Longitudinal-Lateral Coupling: The authors observed that while steering smoothness improved, the acceleration rate distribution for ICODE-MPPI was slightly wider in certain scenarios. This is physically justified by lateral-longitudinal coupling: aggressive steering corrections to maintain lateral safety under heavy disturbances induce lateral tire forces that cause slight losses in forward momentum, requiring more frequent longitudinal acceleration adjustments to maintain reference speed.

5. Significance and Claims

The paper claims that ICODE-MPPI represents a significant advancement in robust path tracking by bridging the gap between idealized modeling and real-world disturbances. By leveraging continuous-time, stability-guaranteed residual learning, the framework:

• Preserves Physical Consistency: Unlike discrete-time learners, ICODEs maintain temporal continuity and physical consistency throughout the MPPI prediction horizon.
• Enhances Robustness: It effectively suppresses control chattering and mitigates steady-state drift, yielding smoother steering commands and superior performance in highly disturbed environments.
• Strategic Trade-offs: The system demonstrates a strategic ability to trade minor transient heading deviations for absolute positional safety, a critical requirement for robust autonomous navigation.

The authors conclude that while the method shows superior performance in simulation, future work will focus on extending the framework to include dynamic obstacle avoidance and validating real-time performance on physical scale-model vehicles.