Robust Path Tracking for Vehicles via Continuous-Time Residual Learning: An ICODE-MPPI Approach
New framework learns unmodeled dynamics to eliminate jittery steering and cut cross-track errors.
Model Predictive Path Integral (MPPI) control is a popular sampling-based strategy for autonomous vehicles, but its performance is limited by the accuracy of its nominal dynamics model. To address this, researchers introduced ICODE-MPPI, a framework that combines MPPI with Input Concomitant Neural Ordinary Differential Equations (ICODEs). These continuous-time learners capture and compensate for unmodeled residual dynamics while preserving physical consistency and temporal continuity over the MPPI prediction horizon, overcoming the limitations of discrete-time approaches.
In high-fidelity simulations of complex trajectories, ICODE-MPPI achieved a 69% reduction in cross-tracking error compared to standard MPPI under persistent disturbances. The method also significantly suppressed control chattering, yielding smoother steering commands and superior robustness. This work, detailed in arXiv paper 2605.03260, promises safer and more reliable autonomous driving in real-world scenarios where model uncertainty is inevitable.
- ICODE-MPPI reduces cross-tracking error by 69% over standard MPPI control under disturbances.
- Uses continuous-time Neural ODEs (ICODEs) to learn unmodeled dynamics while maintaining physical consistency.
- Significantly suppresses control chattering, producing smoother steering commands for better ride quality.
Why It Matters
Enables safer, smoother autonomous vehicle control by compensating for unknown dynamics in real time.