[Paper Review] Baidu Apollo EM Motion Planner

Haoyang Fan, Fan Zhu+8 more|arXiv (Cornell University)|Jul 20, 2018

Autonomous Vehicle Technology and Safety12 references242 citations

TL;DR

This paper presents a real-time, open-source motion planning system (EM planner) for Baidu Apollo that uses a multilane, path-speed iterative approach with DP and spline QP in Frenet space to ensure safety, comfort, and scalability for highway and city driving.

ABSTRACT

In this manuscript, we introduce a real-time motion planning system based on the Baidu Apollo (open source) autonomous driving platform. The developed system aims to address the industrial level-4 motion planning problem while considering safety, comfort and scalability. The system covers multilane and single-lane autonomous driving in a hierarchical manner: (1) The top layer of the system is a multilane strategy that handles lane-change scenarios by comparing lane-level trajectories computed in parallel. (2) Inside the lane-level trajectory generator, it iteratively solves path and speed optimization based on a Frenet frame. (3) For path and speed optimization, a combination of dynamic programming and spline-based quadratic programming is proposed to construct a scalable and easy-to-tune framework to handle traffic rules, obstacle decisions and smoothness simultaneously. The planner is scalable to both highway and lower-speed city driving scenarios. We also demonstrate the algorithm through scenario illustrations and on-road test results. The system described in this manuscript has been deployed to dozens of Baidu Apollo autonomous driving vehicles since Apollo v1.5 was announced in September 2017. As of May 16th, 2018, the system has been tested under 3,380 hours and approximately 68,000 kilometers (42,253 miles) of closed-loop autonomous driving under various urban scenarios. The algorithm described in this manuscript is available at https://github.com/ApolloAuto/apollo/tree/master/modules/planning.

Motivation & Objective

• Address the industrial level-4 motion planning problem with safety, comfort and scalability in a real-time system.
• Develop a multilane framework that handles lane changes via parallel lane-level optimizers and a cross-lane decision mechanism.
• Iteratively optimize path and speed in Frenet coordinates using a dynamic-programming and spline-quadratic-programming (QP) hybrid to respect traffic rules and obstacle decisions.
• Demonstrate scalability to highway and urban driving and provide on-road validation and system deployment metrics.
• Leverage Apollo open-source platform to enable modular, safety-first autonomous driving planning.

Proposed method

• Introduce a multilane framework where candidate lane trajectories are generated in parallel and a cross-lane decider selects the best one based on cost and safety.
• Formulate lane-level optimization as an SL (station-lateral) and ST (station-time) problem using a Frenet frame; perform E-steps (obstacle projection) and M-steps (path and speed optimization).
• Employ a two-stage path optimization: a dynamic-programming (DP) path step to produce a rough tunnel and obstacle decisions, followed by a spline-based quadratic-programming (QP) path refinement within that tunnel.
• Compute speed profiles via a DP-based ST graph search and then refine with a spline QP speed optimization constrained by vehicle dynamics and traffic regulations.
• Solve QP problems efficiently with hot-starts from previous cycles; achieve timing around a few milliseconds per problem, enabling real-time planning (<100 ms on average).
• Balance non-convexity by coupling DP (global-ish search) with QP (convex optimization) for robust path-speed solutions.

Experimental results

Research questions

• RQ1 How can a real-time autonomous motion planner safely handle multilane scenarios with dynamic obstacles while respecting traffic laws?
• RQ2 Can a two-stage DP + spline QP optimization in Frenet space provide scalable, smooth, and safe path and speed profiles for both highway and urban driving?
• RQ3 What is the computational efficiency of the proposed EM planner, and is it feasible for real-time deployment on automotive hardware?
• RQ4 How does the planner manage lane-change decisions (passive vs nonpassive) within a parallel lane framework to maintain safety and ride comfort?
• RQ5 What are the practical deployment metrics and on-road validation results of the EM planner within the Baidu Apollo platform?

Key findings

• The EM planner achieves real-time performance, solving DP+QP path and speed optimization within two-dimensional convex domains and reporting average QP solution times of around 3 ms for the path and speed problems.
• The approach converts a 3D station-lateral-speed problem into two 2D problems, reducing computational complexity and enabling <100 ms planning cycles in practice.
• The system demonstrates deployment across dozens of Apollo vehicles with extensive closed-loop testing, totaling 3,380 hours and about 68,000 km by May 16, 2018.
• The planner handles complex, multi-obstacle urban scenarios and demonstrates a light-decision-based approach that avoids heavyweight rule-based plans while maintaining safety and passability.
• A combination DP + QP mitigates non-convex optimization issues, leveraging DP to explore feasible regions and QP to refine optimal trajectories within convex hulls.