Pylot: A Modular Platform for Exploring Latency-Accuracy Tradeoffs in Autonomous Vehicles (2104.07830v1)

Abstract: We present Pylot, a platform for autonomous vehicle (AV) research and development, built with the goal to allow researchers to study the effects of the latency and accuracy of their models and algorithms on the end-to-end driving behavior of an AV. This is achieved through a modular structure enabled by our high-performance dataflow system that represents AV software pipeline components (object detectors, motion planners, etc.) as a dataflow graph of operators which communicate on data streams using timestamped messages. Pylot readily interfaces with popular AV simulators like CARLA, and is easily deployable to real-world vehicles with minimal code changes. To reduce the burden of developing an entire pipeline for evaluating a single component, Pylot provides several state-of-the-art reference implementations for the various components of an AV pipeline. Using these reference implementations, a Pylot-based AV pipeline is able to drive a real vehicle, and attains a high score on the CARLA Autonomous Driving Challenge. We also present several case studies enabled by Pylot, including evidence of a need for context-dependent components, and per-component time allocation. Pylot is open source, with the code available at https://github.com/erdos-project/pylot.

• The paper presents Pylot, a modular platform that leverages a high-performance dataflow architecture to study latency-accuracy tradeoffs in autonomous vehicle systems.
• It integrates state-of-the-art techniques for object detection, tracking, prediction, and planning, demonstrating seamless transitions between simulation and real-world tests.
• The evaluation highlights practical trade-offs between runtime and accuracy, offering actionable insights for optimizing autonomous vehicle performance.

An Overview of the Pylot Platform for Autonomous Vehicle Research

The paper presents "Pylot," a comprehensive platform designed for advancing research in autonomous vehicle (AV) technology. Developed to facilitate the paper of latency and accuracy trade-offs in AV systems, Pylot embodies a modular approach that leverages a high-performance dataflow architecture. This system allows researchers to explore how various components, such as object detectors and motion planners, influence the end-to-end driving performance of AVs.

Core Contributions and Structure

Pylot's architectural design addresses three primary requirements crucial for AV research: modularity, portability, and debuggability.

1. Modularity: Pylot adopts a dataflow programming model, structuring the AV pipeline as a graph where components are represented as operators. These operators, analogous to ROS nodes, communicate via timestamped message streams. This setup encourages independent innovation by allowing components to be seamlessly swapped and compared, fostering the development of optimal solutions tailored to specific tasks within the AV system.
2. Portability: The platform is built atop a high-throughput dataflow system that supports seamless transitions between simulator environments, like CARLA, and real-world AVs, such as the Lincoln MKZ. This capability is bolstered by a synchronizer for simulation that emulates real-world component runtimes, thus ensuring consistency between simulated and actual driving scenarios.
3. Debuggability: By integrating with CARLA’s ScenarioRunner, Pylot facilitates deterministic execution and replay of driving scenarios, crucial for debugging and testing. The system logs extensive data regarding runtimes and module outputs, enabling comprehensive analyses of AV behavior in various contexts.

Reference Implementations

Pylot includes several state-of-the-art implementations for each major AV module:

• Object Detection: Utilizes models such as Faster-RCNN and EfficientDet to address runtime-accuracy trade-offs.
• Object Tracking: Offers different tracking approaches, including SORT and DeepSORT, to maintain consistent object identifiers across frames.
• Prediction: Implements various ML-driven methods like R2P2 and Multipath for forecasting agent trajectories.
• Planning and Control: Features multiple planning strategies such as RRT* and Hybrid A* for path optimization, alongside control algorithms like PID and MPC for vehicle maneuvering.

These components, tested both in simulation and on a real vehicle, provide a comprehensive testing ground for evaluating component interactions and overall system performance.

Evaluation and Case Studies

The paper explores Pylot's capabilities through illustrative case studies that examine:

• Runtime-Accuracy Trade-offs: The introduction of "timely accuracies" as metrics demonstrates how runtime impacts module accuracy within dynamic environments, revealing critical insights for balancing trade-offs.
• End-to-end Performance: By simulating emergency scenarios, Pylot enables the evaluation of different planning configurations, illustrating their distinct impacts on AV safety and passenger comfort.

Implications and Future Directions

Practically, Pylot serves as a robust platform for AV R&D, providing researchers with a versatile toolset for testing and refining technologies under realistic conditions. Theoretically, it underscores the importance of modularity and portability in AV development, setting a precedent for future platforms in this field.

The authors propose future enhancements, including the integration of reinforcement learning models, which could further broaden the scope of AV research facilitated by Pylot.

In conclusion, Pylot represents a significant contribution to AV research, providing a scalable and flexible platform for exploring the intricacies of autonomous driving systems. Its open-source nature invites collaboration and innovation, driving the autonomous vehicle field toward more reliable and efficient solutions.