[Paper Review] TDR-OBCA: A Reliable Planner for Autonomous Driving in Free-Space Environment

Runxin He, Jinyun Zhou+7 more|arXiv (Cornell University)|Sep 23, 2020

Robotic Path Planning Algorithms20 references4 citations

TL;DR

TDR-OBCA is an optimization-based trajectory planner for autonomous driving in free-space environments, enhancing robustness, driving comfort, and efficiency through temporal and dual variable warm starts and problem reformulation. It reduces failure rates by 96.67%, improves driving comfort by 13.53%, and increases planner efficiency by up to 44.82% across varying obstacle counts, validated in simulations and hundreds of hours of real-world testing in the U.S. and China.

ABSTRACT

This paper presents an optimization-based collision avoidance trajectory generation method for autonomous driving in free-space environments, with enhanced robustness, driving comfort and efficiency. Starting from the hybrid optimization-based framework, we introduces two warm start methods, temporal and dual variable warm starts, to improve the efficiency. We also reformulate the problem to improve the robustness and efficiency. We name this new algorithm TDR-OBCA. With these changes, compared with original hybrid optimization we achieve a 96.67% failure rate decrease with respect to initial conditions, 13.53% increase in driving comforts and 3.33% to 44.82% increase in planner efficiency as obstacles number scales. We validate our results in hundreds of simulation scenarios and hundreds of hours of public road tests in both U.S. and China. Our source code is available at https://github.com/ApolloAuto/apollo.

Motivation & Objective

• To address the high failure rate and inefficiency of existing trajectory planners in free-space autonomous driving scenarios such as parking and hailing.
• To improve robustness against diverse initial conditions and obstacle configurations in nonholonomic, kinodynamically constrained environments.
• To enhance driving comfort by minimizing abrupt steering commands through penalized control inputs in the optimization cost function.
• To increase computational efficiency for real-time deployment, especially as obstacle count increases.
• To validate the planner’s performance across diverse real-world conditions through extensive simulation and on-road testing in both the U.S. and China.

Proposed method

• Introduces temporal warm start by reusing solution trajectories from the previous time step to accelerate convergence.
• Applies dual variable warm start by initializing dual variables from the previous solution, improving dual convergence in sequential quadratic programming.
• Reformulates the collision avoidance constraints using dual expressions to eliminate binary variables and convert the problem into a continuous, large-scale Model Predictive Control (MPC) problem.
• Incorporates smoothness penalties on steering and acceleration inputs to enhance driving comfort and reduce control effort.
• Integrates TDR-OBCA into the Apollo Autonomous Driving Platform, where it is triggered when the ego vehicle is off-road or in free-space maneuvers.
• Uses trajectory stitching, safety checks, and gear-based partitioning to prepare the output for control execution via CAN bus.

Experimental results

Research questions

• RQ1 Can warm start techniques significantly reduce the failure rate of trajectory planners in free-space autonomous driving under varying initial conditions?
• RQ2 To what extent can reformulating collision constraints into dual forms improve computational efficiency and robustness in non-convex MPC problems?
• RQ3 How does the inclusion of smoothness penalties in the cost function affect driving comfort and control effort in real-world scenarios?
• RQ4 Can the proposed method maintain real-time performance as the number of obstacles increases, especially in complex environments with moving pedestrians and vehicles?
• RQ5 Does the TDR-OBCA algorithm demonstrate consistent performance across diverse real-world driving conditions, including parking, hailing, and three-point turns?

Key findings

• The failure rate of trajectory planning dropped from 37.5% to 1.25% across different initial conditions, representing a 96.67% reduction in failure rate.
• Driving comfort improved by over 13.53% as measured by reduced steering control output variance in the optimization solution.
• Planner efficiency improved by 3.33% to 44.82% across scenarios with increasing obstacle counts, including moving obstacles and boundary constraints.
• In real-world road tests across the U.S. and China, lateral tracking errors remained within 0.01–0.2 meters, and heading errors were below 0.14 degrees at final poses.
• The algorithm successfully generated smooth, kinematically feasible trajectories for complex maneuvers such as reverse parking, short reverse parking, and zig-zag parking with high accuracy.
• The method demonstrated scalability and robustness in both simulation (282 cases) and hundreds of hours of real-world testing, confirming its viability for deployment in production autonomous driving systems.