High-speed autonomy in the real world requires accurate control, which often relies on dynamics models that capture the complex interaction between a robot and its environment. In off-road regimes, this terrain interaction dominates the robot's dynamics, driven by challenging characteristics such as diverse surface properties, complex geometries, environmental diversity, and high-speed instability. Physics-based dynamics models lack the expressivity for these terradynamics, and existing learned models do not fully exploit available perception. This thesis investigates how terrain-aware perception can improve learned dynamics modeling and control at high speed and how such models can be rigorously evaluated before deployment.

First, we demonstrate how perception representations can make dynamics models terrain-aware, capturing the geometric and semantic detail that drives terrain interaction. We formulate a representation that queries terrain features along the robot's predicted motion, yielding higher prediction accuracy. Second, we introduce a rigorous evaluation method to expose failure modes before real-world deployment. We collect a challenging, multi-season dataset at speeds up to 13 m/s and mine the most difficult evaluation samples using our benchmarking method. We observe that, while models appear accurate on average, our benchmark surfaces the long-tail failure cases where prior models fail catastrophically.

Together with our verified, terrain-aware model, we decrease the worst-case prediction error by 23.8%, compared to physics-based and learned baselines. We further evaluate on a full-scale ATV platform across high-speed (≥10m/s) and geometrically challenging courses with a 34.9% reduction in maximum cross-track error. These results demonstrate the importance of embedding environment context for locomotion-related tasks.