PLUTO is a novel framework that advances imitation learning-based planning for autonomous driving. The framework addresses three key challenges: (1) multi-modal driving behavior generation, (2) limitations of pure imitation learning, and (3) causal understanding. The model architecture enables flexible and diverse driving behaviors by integrating longitudinal and lateral queries. A novel auxiliary loss computation method based on differentiable interpolation allows efficient batch-wise computation. Additionally, a contrastive imitation learning (CIL) framework is introduced, incorporating new data augmentations to regulate driving behaviors and enhance interaction learning. PLUTO achieves state-of-the-art closed-loop performance on the nuPlan dataset, surpassing the current top-performing rule-based planner. The framework is evaluated using the nuPlan benchmark, demonstrating superior performance in closed-loop planning. PLUTO's contributions include a query-based model architecture, a novel auxiliary loss calculation method, and the CIL framework. The model and benchmark are publicly available.PLUTO is a novel framework that advances imitation learning-based planning for autonomous driving. The framework addresses three key challenges: (1) multi-modal driving behavior generation, (2) limitations of pure imitation learning, and (3) causal understanding. The model architecture enables flexible and diverse driving behaviors by integrating longitudinal and lateral queries. A novel auxiliary loss computation method based on differentiable interpolation allows efficient batch-wise computation. Additionally, a contrastive imitation learning (CIL) framework is introduced, incorporating new data augmentations to regulate driving behaviors and enhance interaction learning. PLUTO achieves state-of-the-art closed-loop performance on the nuPlan dataset, surpassing the current top-performing rule-based planner. The framework is evaluated using the nuPlan benchmark, demonstrating superior performance in closed-loop planning. PLUTO's contributions include a query-based model architecture, a novel auxiliary loss calculation method, and the CIL framework. The model and benchmark are publicly available.