Overview
Mosaic is a structured decision-making framework designed to integrate rule-based and learned motion planners in autonomous driving applications. It differentiates trajectory verification and selection from the trajectory generation processes of individual planners. This framework utilizes arbitration graphs to facilitate the transparent and traceable decision-making process. The system aims to combine the strengths of both rule-based and learned planning paradigms to address challenges in safety, explainability, adaptability, and resilience to real-world traffic complexity.
Research Context
Autonomous driving systems rely on motion planners to navigate complex environments safely and effectively. Existing approaches often fall into two main categories: rule-based planners and learned planners. Rule-based planners are characterized by their predictable and explainable behavior, which contributes to safety and transparency. However, these planners frequently encounter difficulties in interpreting the intricacies and uncertainties of real-world traffic scenarios. In contrast, learned planners demonstrate robust adaptability to diverse conditions but are often limited by reduced transparency, making their decisions harder to interpret. They can also exhibit occasional safety violations. The research addresses the challenge of creating safe and explainable motion planning systems that can handle the complexity and uncertainty inherent in autonomous driving, while mitigating the individual shortcomings of standalone planning approaches.
Approach
Mosaic implements its integration through arbitration graphs, which serve to decouple the processes of trajectory verification and selection from the generation of trajectories by individual planners. This separation allows for independent contributions from verification and trajectory selection. Centralized verification functions as a safety floor, providing a foundational layer of safety assurance. Per-step trajectory selection operates as a performance ceiling, enabling the system to leverage the complementary strengths of both rule-based and learned planners. The framework is designed to ensure that each decision made is transparent and traceable within the system.
Findings
- Mosaic's centralized verification mechanism reduced at-fault collisions from 25 for each standalone planner to 16.
- In experimental evaluation on the nuPlan Val14 closed-loop benchmark, Mosaic achieved 95.56 CLS-NR and 94.18 CLS-R. These scores represent a new state of the art on this benchmark.
- On the interPlan benchmark, which focuses on highly interactive and out-of-distribution scenarios, Mosaic scored 54.10 CLS-R.
- Mosaic outperformed its best constituent planner on the interPlan benchmark by 22.8% without requiring retraining or additional data.
Why This Matters
The integration of rule-based and learned planning paradigms within Mosaic directly addresses central challenges in autonomous driving, specifically regarding the need for both safety and explainability in real-world traffic conditions. By separating trajectory verification and selection from generation, the framework provides a structured approach that aims to enhance both predictability and adaptability. The observed reduction in at-fault collisions and improved performance on specified benchmarks suggest a potential for more robust and reliable autonomous driving systems. This approach allows for a clearer understanding of decision-making processes, which is relevant for the development and deployment of autonomous vehicles.
Potential Applications
The Mosaic framework is developed for applications in autonomous driving motion planning, as evidenced by its evaluation on nuPlan and interPlan benchmarks. Its ability to integrate different planning paradigms could be applied in scenarios requiring both predictable, explainable behavior and adaptive performance in complex and uncertain traffic conditions.
Key Limitations Mentioned by Researchers
No specific limitations are explicitly mentioned in the provided source material.