This paper introduces Terrain-Attentive Learning (TAL), a 6-DoF kinodynamics learning approach that focuses on efficiently modeling vehicle-terrain interactions on vertically challenging terrain. Traditional kinodynamic models for wheeled robots assume 2D motion and are limited in their ability to handle complex, 3D terrain interactions. TAL addresses this by learning a model that is attentive to the specific terrain features critical to the current vehicle-terrain interaction, enabling efficient real-time querying in onboard motion planners. The model is trained using a dataset of vehicle states, control inputs, and terrain information, and is combined with a state-of-the-art sampling-based motion planner to enable efficient kinodynamic planning. TAL uses representation learning to extract relevant terrain features from elevation maps, allowing the model to focus on the terrain that is most relevant to the current vehicle state. This approach significantly reduces model prediction error compared to a state-of-the-art kinodynamic model for vertically challenging terrain, achieving an average reduction of 51.1% in model prediction error across all six degrees of freedom. The model is implemented on a 1/10th-scale unmanned ground vehicle (V4W) and tested in a rock testbed with varying terrain configurations. The results show that TAL can accurately predict future vehicle states and enable efficient navigation through vertically challenging terrain. The TAL model is implemented using a combination of convolutional and multi-layer perceptron networks to extract terrain features and predict future vehicle states. The model is trained using a self-supervised representation loss that ensures the learned terrain features are sufficient for accurate prediction. The model is then used in a Model Predictive Path Integral (MPPI) planner to generate feasible motion plans for navigation through vertically challenging terrain. The results show that TAL outperforms other approaches in terms of prediction accuracy and enables more efficient navigation through complex terrain. The approach demonstrates the potential of data-driven kinodynamic modeling for autonomous navigation on vertically challenging terrain.This paper introduces Terrain-Attentive Learning (TAL), a 6-DoF kinodynamics learning approach that focuses on efficiently modeling vehicle-terrain interactions on vertically challenging terrain. Traditional kinodynamic models for wheeled robots assume 2D motion and are limited in their ability to handle complex, 3D terrain interactions. TAL addresses this by learning a model that is attentive to the specific terrain features critical to the current vehicle-terrain interaction, enabling efficient real-time querying in onboard motion planners. The model is trained using a dataset of vehicle states, control inputs, and terrain information, and is combined with a state-of-the-art sampling-based motion planner to enable efficient kinodynamic planning. TAL uses representation learning to extract relevant terrain features from elevation maps, allowing the model to focus on the terrain that is most relevant to the current vehicle state. This approach significantly reduces model prediction error compared to a state-of-the-art kinodynamic model for vertically challenging terrain, achieving an average reduction of 51.1% in model prediction error across all six degrees of freedom. The model is implemented on a 1/10th-scale unmanned ground vehicle (V4W) and tested in a rock testbed with varying terrain configurations. The results show that TAL can accurately predict future vehicle states and enable efficient navigation through vertically challenging terrain. The TAL model is implemented using a combination of convolutional and multi-layer perceptron networks to extract terrain features and predict future vehicle states. The model is trained using a self-supervised representation loss that ensures the learned terrain features are sufficient for accurate prediction. The model is then used in a Model Predictive Path Integral (MPPI) planner to generate feasible motion plans for navigation through vertically challenging terrain. The results show that TAL outperforms other approaches in terms of prediction accuracy and enables more efficient navigation through complex terrain. The approach demonstrates the potential of data-driven kinodynamic modeling for autonomous navigation on vertically challenging terrain.