Physics-Constrained Imitation Learning for Autonomous Racing
Haohan Yang, Haochen Liu, Zhou Yanxin, Shuge Wu, Chen Lv
AI summary
Problem
Traditional control methods are highly sensitive to modeling parameters, while end-to-end learning approaches suffer from distributional shifts and unpredictable failures when operating at a vehicle's dynamic limits.
Approach
The PCIL framework pairs a lightweight imitation policy with a physics-based dynamic model that automatically triggers a conservative fallback controller when physical constraints are breached, iteratively retraining on the collected fallback data.
Key results
- 100% simulation success rate eliminating irreversible collisions
- Competitive lap time of 109.82 seconds on Yas Marina circuit
- Progressive performance gains via iterative fallback data augmentation
- Superior safety and reliability over state-of-the-art RL and IL baselines
Why it matters
Provides a safe, high-performance framework for extreme-condition autonomous driving that reliably bridges learning-based adaptability with physics-based guarantees.
Abstract
Autonomous racing has become increasingly pop- ular in both academia and industry as a testbed for pushing general autonomous driving modules, such as perception, plan- ning, and control, to their limits. Although traditional control approaches can generate optimal control sequences at the edge of the racing vehicles’ physical controllability, they are highly sensitive to the accuracy of modeling parameters, such as tire model coefficients. Meanwhile, end-to-end learning methods are susceptible to distributional shifts, leading to unpredictable and irreversible failures. To address these challenges, this work introduces a physics-constrained imitation learning (PCIL) framework that effectively leverages the advantages of deep learning techniques and knowledge-driven strategies. Specif- ically, a fallback strategy would be automatically triggered when the vehicle states exceed predefined physical constraints. Meanwhile, the data from the knowledge-driven strategy will be augmented into the original dataset, and repeated re-training using an aggregated dataset could progressively improve PCIL. A series of simulations and real-world shadow testing are conducted at the Yas Marina circuit, and experimental results demonstrate superior performance compared to state-of-the-art methods, which suggests that it provides a promising solution for real-world autonomous racing.