Curriculum Reinforcement Learning for Quadrotor Racing with Random Obstacles
Fangyu Sun, Fanxing Li, Yu Hu, Linzuo Zhang, Yueqian Liu, Wenxian Yu, Danping Zou
AI summary
Problem
Existing drone racing methods focus on obstacle-free tracks and struggle to balance high-speed gate traversal with obstacle avoidance, leading to poor generalization and low real-world success rates.
Approach
The authors propose a vision-based curriculum reinforcement learning framework that progressively increases training difficulty, uses domain randomization, and employs a multi-scene updating strategy to train an end-to-end policy that balances collision avoidance and gate passing.
Key results
- Multi-stage curriculum learning enables systematic generalization to unseen cluttered environments.
- Multi-scene updating paradigm significantly improves RL training efficiency and convergence.
- Custom reward function resolves the inherent conflict between obstacle avoidance and gate traversal.
- Achieves 100% success rate and speeds up to 10 m/s in simulation and 8 m/s in real-world flights.
Why it matters
Provides a robust, deployable solution for high-speed autonomous navigation in complex, real-world cluttered environments, advancing practical agile robotics applications.
Abstract
Autonomous drone racing has attracted increasing interest as a research topic for exploring the limits of agile flight. However, existing studies primarily focus on obstacle- free racetracks, while the perception and dynamic challenges introduced by obstacles remain underexplored, often result- ing in low success rates and limited robustness in real- world flight. To this end, we propose a novel vision-based curriculum reinforcement learning framework for training a robust controller capable of addressing unseen obstacles in drone racing. We combine multi-stage cu rriculum learning, domain randomization, and a multi-scene updating strategy to address the conflicting challenges of obstacle avoidance and gate traversal. Our end-to-end control policy is implemented as a single network, allowing high-speed flight of quadrotors in environments with variable obstacles. Both hardware-in-the- loop and real-world experiments demonstrate that our method achieves faster lap times and higher success rates than existing approaches, effectively advancing drone racing in obstacle-rich environments. The video and code are available at: https:// github.com/SJTU-ViSYS-team/CRL-Drone-Racing.