Meta-Learning Online Dynamics Model Adaptation in Off-Road Autonomous Driving

TL;DR

This work tackles the challenge of high-speed off-road autonomy where terrain changes can drastically alter vehicle dynamics. It introduces a meta-learning framework that couples a Kalman-filter–based online adaptation of a physics-informed dynamics model with offline optimization of adaptation basis functions, enabling real-time, robust updates. The approach demonstrates improved prediction accuracy, faster and safer closed-loop behavior, and successful validation on a full-scale vehicle as well as simulated environments, outperforming non-adaptive and non-meta-learned baselines. By enabling rapid, terrain-aware adaptation within a model-predictive control context, the method advances reliable navigation across diverse and unseen environments and can be extended to a broader class of model-based controllers.

Abstract

High-speed off-road autonomous driving presents unique challenges due to complex, evolving terrain characteristics and the difficulty of accurately modeling terrain-vehicle interactions. While dynamics models used in model-based control can be learned from real-world data, they often struggle to generalize to unseen terrain, making real-time adaptation essential. We propose a novel framework that combines a Kalman filter-based online adaptation scheme with meta-learned parameters to address these challenges. Offline meta-learning optimizes the basis functions along which adaptation occurs, as well as the adaptation parameters, while online adaptation dynamically adjusts the onboard dynamics model in real time for model-based control. We validate our approach through extensive experiments, including real-world testing on a full-scale autonomous off-road vehicle, demonstrating that our method outperforms baseline approaches in prediction accuracy, performance, and safety metrics, particularly in safety-critical scenarios. Our results underscore the effectiveness of meta-learned dynamics model adaptation, advancing the development of reliable autonomous systems capable of navigating diverse and unseen environments. Video is available at: https://youtu.be/cCKHHrDRQEA

Figures (6)

• Figure 1: Trajectories for a single 3-lap run, with insets displaying video stills. The baseline configuration shows erratic trajectories with frequent course boundary and rollover limit violations. In contrast, our adaptation configuration demonstrates more deliberate and compliant trajectories as the car learns the terrain dynamics in real time.
• Figure 2: Meta-learning online dynamics model adaptation. Online, a Kalman filter updates the linear combination of an ensemble of last-layer weights (\ref{['alg:adapt']}). Offline, trajectory segments are used to meta-learn model parameters, the last-layer ensemble, and the Kalman filter parameters (\ref{['alg:meta']}).
• Figure 3: Forward facing camera stills from the dataset highlighting a diverse range of terrains: A) flat sandy beach with a mixture of packed wet sand and loose dry sand; B) wet dense mud that forms deep ruts; C) dirt trails with low dry grass that weave through dense trees; D) mixed vegetation including dry, dense vehicle-height grass; E) dense overgrown mixed vegetation ranging in crushability; and F) loose gravel, uneven ground, and steep slopes.
• Figure 4: Trajectories for all 3-lap real-world runs. The baseline configuration exhibits erratic motion, frequently violating course boundaries and rollover limits. In contrast, our adaptation configuration produces more deliberate and compliant trajectories, as the vehicle learns the terrain dynamics in real time.
• Figure 5: Norm of the adapted parameters during one of the real-world runs.

Theorems & Definitions (6)

• Remark 1
• Remark 2
• Remark 3
• Remark 4
• Remark 5
• Remark 6