Head-to-Head autonomous racing at the limits of handling in the A2RL challenge

TL;DR

This work presents a fully integrated autonomous racing stack from TUM that operates at the handling limits on a real Formula1–style track. It combines offline raceline optimization and a GripMap of spatially resolved dynamic constraints with online perception, localization, prediction, and a sampling-based local planner under a Tube MPC–based control loop. Key contributions include a modular real-world software architecture, a grip-aware planning framework, a robust multi-sensor localization pipeline capable of GNSS-denied operation, and a comprehensive deployment/testing strategy validated by a real-world A2RL victory. The results demonstrate near-human performance (within 10% of a professional driver) and highlight practical insights and limitations—particularly the need for better opponent modeling and real-time grip-map learning—to push autonomous racing closer to the absolute performance frontier.

Abstract

Autonomous racing presents a complex challenge involving multi-agent interactions between vehicles operating at the limit of performance and dynamics. As such, it provides a valuable research and testing environment for advancing autonomous driving technology and improving road safety. This article presents the algorithms and deployment strategies developed by the TUM Autonomous Motorsport team for the inaugural Abu Dhabi Autonomous Racing League (A2RL). We showcase how our software emulates human driving behavior, pushing the limits of vehicle handling and multi-vehicle interactions to win the A2RL. Finally, we highlight the key enablers of our success and share our most significant learnings.

Figures (8)

• Figure 1: The EAV24 race car
• Figure 2: Software architecture of TUM Autonomous Motorsport. Blue components run on the vehicle during operation, orange components are executed offline, and white components run on a remote machine. For simplification, an orange arrow indicates modules that subscribe to the current ego state.
• Figure 3: Comparison of the point cloud before and after pre-processing. The ground truth annotation is visualized as a red box.
• Figure 4: Illustration of the used static point cloud map. The coloring of the drawn trajectory indicates the GNSS standard deviation (STDEV) on the different parts of the race track.
• Figure 5: Time-optimal raceline through turns 6 and 7. To improve visibility, the controller induced lateral offset to the racing line is increased by a factor of 2 in this figure compared to the real world. The grayscale grid shows the spatially resolved acceleration limits used for optimization and online planning.