WANDR (Wrist-driven Autonomous Navigation for Data-based goal Reaching) is a novel data-driven model designed to generate realistic and natural human motions that can reach arbitrary 3D goals. The model uses a conditional Variational Autoencoder (c-VAE) to condition the generation process on the initial pose and the goal position, guided by *intention* features that steer the human's orientation, position, and wrist towards the goal. These intention features are derived from the current pose and the goal location, allowing the model to adapt to novel situations without predefined sub-goals or motion paths. WANDR is trained on the AMASS and CIRCLE datasets, which capture a wide range of locomotion and goal-reaching motions, respectively. The method is evaluated extensively, demonstrating its ability to generate long-term and natural motions that reach 3D goals and generalize to unseen goal locations. The results highlight the effectiveness of the intention mechanism in guiding the motion generation process and enabling the incorporation of pseudo-goal labels for datasets lacking explicit goal annotations. The model and code are available for research purposes.WANDR (Wrist-driven Autonomous Navigation for Data-based goal Reaching) is a novel data-driven model designed to generate realistic and natural human motions that can reach arbitrary 3D goals. The model uses a conditional Variational Autoencoder (c-VAE) to condition the generation process on the initial pose and the goal position, guided by *intention* features that steer the human's orientation, position, and wrist towards the goal. These intention features are derived from the current pose and the goal location, allowing the model to adapt to novel situations without predefined sub-goals or motion paths. WANDR is trained on the AMASS and CIRCLE datasets, which capture a wide range of locomotion and goal-reaching motions, respectively. The method is evaluated extensively, demonstrating its ability to generate long-term and natural motions that reach 3D goals and generalize to unseen goal locations. The results highlight the effectiveness of the intention mechanism in guiding the motion generation process and enabling the incorporation of pseudo-goal labels for datasets lacking explicit goal annotations. The model and code are available for research purposes.