Qian, Zelin (Tsinghua University), Jiang, Kun (Tsinghua University), Cao, Zhong (University of Michigan), Qian, Kangan (Tsinghua University), Xu, Yiliang (Zongmu Technology), Zhou, Weitao (Tsinghua University), Yang, Diange (State Key Laboratory of Automotive Safety and Energy, Collaborat)

SPIDER: Self-Driving Planners and Intelligent Decision-Making Engines with Reusability

Scheduled for presentation during the Invited Session "Driving the Edge: Addressing Corner Cases in Self-driving Vehicles" (WeBT5), Wednesday, September 25, 2024, 15:50−16:10, Salon 13

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

The rapid development of self-driving technology owes much to the growth of open-source algorithms, which facilitate the rapid reproduction of existing solutions and foster innovation. However, in the realm of planning and decision-making, the diversity of technical approaches, e.g., rule-based and learning-based planners, often restricts knowledge sharing, as open-source algorithms typically cater to specific methodologies. Additionally, developing and testing a planner from scratch from scratch is still a very complex process involving multiple orthogonal steps. To address these challenges, this paper introduces SPIDER, a tool-chain designed to enhance knowledge sharing and reuse across different planning technology approaches. SPIDER offers a common and reusable suite of basic planning tools and supports the entire lifecycle of planner development, from initial design to deployment. Key features of SPIDER include the standardization of interfaces within the planning modules and the integration of foundational planning frameworks. The toolchain provides a variety of tools and modules commonly utilized by both rule-based and learning-based planners and establishes a data closed-loop engine that supports both imitation learning (IL) and reinforcement learning (RL). Additionally, SPIDER includes straightforward testing tools and interfaces, ensuring compatibility with mainstream simulators and datasets. The installation of SPIDER is streamlined, requiring only a single command, and allows for the rapid implementation and testing of planner examples with just a few lines of code. The unified interface framework of SPIDER also facilitates swift migration and deployment across diverse testing environments. SPIDER is open source and available at https://github.com/Thu-ADLab/SPIDER.