TL;DR
This paper introduces µDRIVE, a novel Domain-Specific Language (DSL) that allows users to directly control the planning behavior of autonomous driving systems (ADSs) by specifying their preferences through event-driven rules. This approach aims to enhance passenger comfort and safety by enabling personalized driving experiences and improving compliance with traffic regulations, as demonstrated by successful integration with the Apollo ADS framework.
What Problem Does This Paper Address?
Autonomous vehicles (AVs) rely on complex Autonomous Driving Systems (ADSs) to navigate and drive. While these systems are becoming increasingly sophisticated, they often lack a mechanism to systematically incorporate the individual preferences and intentions of passengers into their decision-making processes. This can lead to a less satisfying and potentially less comfortable ride, as the AV might not drive in a way that aligns with the user's expectations or comfort level. Furthermore, the "black box" nature of some ADSs can lead to a perceived lack of human control, which is a significant concern for public acceptance. Existing approaches often treat driver input as mere auxiliary information for perception or navigation, rather than a direct means to influence the core driving plan. This paper addresses the gap of directly integrating user preferences into the ADS planning module to create a more personalized and user-centric autonomous driving experience.
Methods
The core contribution of this paper is µDRIVE, an event-based Domain-Specific Language (DSL) designed for specifying autonomous vehicle behavior. µDRIVE allows users to define rules that are triggered by specific events occurring in the AV's environment (e.g., detecting an obstacle, entering a specific road type). These rules can include conditions that must be met for the rule to activate and actions that dynamically adjust parameters within the ADS planning module. This approach allows users to express preferences for driving style (e.g., speed, lane changes, acceleration) and adapt them in real-time based on the driving situation. The language is designed to be high-level, avoiding low-level control commands, and to integrate seamlessly with existing ADS frameworks. The authors implemented µDRIVE using the Antlr4 parser and integrated it with the Apollo ADS framework (version 9.0). The system comprises a language parser and a language interface that translates µDRIVE commands into parameter adjustments for Apollo's planning module.
Datasets
The paper does not explicitly mention the use of datasets for training µDRIVE itself. However, for evaluation, they utilized a benchmark of violation-inducing scenarios for Chinese traffic laws, referencing the "Lawbreaker" dataset/framework [48]. This dataset is used to assess how µDRIVE intervention impacts the ADS's compliance with traffic regulations.
Key Insights and Ideas
- Direct Control over Planning: The most significant insight is the shift from treating driver input as auxiliary information to enabling direct intervention in the ADS's planning module. This allows for a more profound level of user control.
- Event-Driven Personalization: µDRIVE's event-based rule system allows for dynamic and context-aware personalization of the driving experience, adapting to real-time environmental changes.
- Bridging the User-ADS Gap: The language provides a high-level abstraction that empowers users to express their driving intentions without needing to understand the intricacies of the underlying ADS.
- Maintaining ADS Integrity: The approach focuses on adjusting planning parameters rather than modifying core ADS modules, preserving the integrity and modularity of existing systems like Apollo.
- User Confidence and Comfort: By allowing users to influence driving behavior, µDRIVE has the potential to significantly increase passenger confidence and comfort in AVs.
Key Results
- Effective Control: Experiments demonstrated that users can effectively influence Apollo's planning through µDRIVE, leading to improved compliance with traffic regulations.
- Traffic Law Compliance: In scenarios designed to test compliance with Chinese traffic laws, µDRIVE intervention helped the AV achieve 100% compliance in feasible cases, significantly outperforming the original Apollo system (which often had success rates below 50%).
- Low Latency: The response time for µDRIVE commands was consistently in the millisecond or sub-millisecond level, ensuring real-time effectiveness. Parsing time averaged 0.98 ms per action/condition, and actions generally took effect within about 100 ms.
- Manoeuvre-Specific Delays: Certain complex maneuvers like re-planning and lane changing took approximately 1 second due to the underlying Apollo system's processing requirements.
Limitations and Future Work
- Platform Dependence (for specific features): While µDRIVE is designed to be platform-agnostic, the integration with Apollo meant that some features (e.g., lights, horns) were not supported due to Apollo's current interfaces. Adapting these interfaces is considered a straightforward process.
- Simulator Limitations: The authors encountered limitations with simulators, preventing the testing of certain traffic laws (e.g., fog lights compliance).
- Future Work:
- LLM-based µDRIVE Generation: Exploring the use of Large Language Models (LLMs) to automatically generate µDRIVE rules from driver manuals.
- Data-Driven Personalization: Synthesizing customized µDRIVE rules based on a user's driving data to enhance adaptability and personalization.
Summary updated