Autonomous Navigation
Vision-based UAV obstacle avoidance with deep reinforcement learning, imitation learning, AirSim simulation, and real-flight validation.
This project developed a learning-based local planner for UAV visual navigation and obstacle avoidance in unknown environments.
The work combines deep reinforcement learning, imitation learning, and simulation-to-real validation. A training platform was built on AirSim with OpenAI Gym-style interfaces, kinematic models for multirotor and fixed-wing UAVs, and multiple Unreal Engine environments for policy training and testing.
Key contributions:
- Built an open-source AirSim-based UAV navigation framework with 550+ GitHub stars and 70+ forks.
- Implemented multirotor, fixed-wing, and flapping-wing kinematic simulation environments.
- Trained local obstacle-avoidance policies in simulation and transferred them to real flight tests.
- Developed a training-state visualization interface for monitoring learning progress.
- Studied model interpretability using Shapley-value-based output contribution analysis, action-level explanation, and feature-correlation analysis.
Autonomous flight with a trained neural network policy in simulation.