This paper addresses the challenge of wheeled robots navigating vertically challenging terrain, where the vehicle's pose in all six Degrees-of-Freedom (DoFs) can vary significantly due to imbalanced contact forces, varying momentum, and chassis deformation. To efficiently model and predict these complex interactions, the authors propose Terrain-Attentive Learning (TAL), a 6-DoF kinodynamics learning approach that focuses on the specific underlying terrain critical to the current vehicle-terrain interaction. TAL is designed to be queried in real-time motion planners onboard small robots, significantly reducing model prediction errors compared to state-of-the-art models. The approach combines a state-action encoder and a kinodynamics predictor, leveraging self-supervised representation learning to efficiently process robot perception data and predict future vehicle states. Experimental results on a 1/10th-scale unmanned ground vehicle demonstrate that TAL achieves an average 51.1% reduction in model prediction error among all 6 DoFs, outperforming other models in both accuracy and efficiency. The paper also discusses the trade-offs between model fidelity and planning frequency, highlighting the need for further research in this area.This paper addresses the challenge of wheeled robots navigating vertically challenging terrain, where the vehicle's pose in all six Degrees-of-Freedom (DoFs) can vary significantly due to imbalanced contact forces, varying momentum, and chassis deformation. To efficiently model and predict these complex interactions, the authors propose Terrain-Attentive Learning (TAL), a 6-DoF kinodynamics learning approach that focuses on the specific underlying terrain critical to the current vehicle-terrain interaction. TAL is designed to be queried in real-time motion planners onboard small robots, significantly reducing model prediction errors compared to state-of-the-art models. The approach combines a state-action encoder and a kinodynamics predictor, leveraging self-supervised representation learning to efficiently process robot perception data and predict future vehicle states. Experimental results on a 1/10th-scale unmanned ground vehicle demonstrate that TAL achieves an average 51.1% reduction in model prediction error among all 6 DoFs, outperforming other models in both accuracy and efficiency. The paper also discusses the trade-offs between model fidelity and planning frequency, highlighting the need for further research in this area.