PLUTO: Pushing the Limits of Imitation Learning for Autonomous Driving

PLUTO introduces a sophisticated architecture that combines a longitudinal-lateral aware model, an innovative auxiliary loss computation, and a novel training framework augmenting contrastive learning. This comprehensive approach enables PLUTO to surpass existing rule-based and learning-based methods in autonomous driving. Noteworthy aspects include:

• Enhanced model architecture that promotes diverse driving behaviors
• Efficient batch-wise computation of auxiliary loss
• Regulatory data augmentations influenced by real-world driving datasets
• Utilization of the large-scale real-world nuPlan dataset and its associated standardized planning benchmark

PLUTO’s state-of-the-art performance is documented with extensive tests, showing superiority over prior approaches and providing practical implementations for autonomous driving systems. The framework not only represents a significant stride towards understanding vehicle interactions but also sets new benchmarks for closed-loop performance in autonomous vehicles. The available results and codebase offer valuable resources for further exploration and adaptation in the field.

Given its exceptional advancements, PLUTO could serve as a model for future developments in the sector, hinting at broader applications beyond standard environments, potentially involving more dynamic and unpredicted scenarios.