Imitation Learning for Unmanned Aerial Vehicle Obstacle Avoidance Based on Visual Features with DAgger

Unmanned aerial vehicles (UAVs) face the challenge of autonomous obstacle avoidance in complex, multi-obstacle environments. Behavior cloning offers a promising approach to rapidly acquire a learning policy from limited expert demonstrations. However, pure imitation learning inherently suffers from poor exploration and limited generalization, typically necessitating extensive datasets to train competent student policies. We utilize a cross-modal variational autoencoder (CM-VAE) to extract compact features from raw visual inputs and UAV states, which then feed into a policy network. We evaluated our approach in a simulated environment featuring a challenging circular trajectory with eight gate obstacles. The results demonstrate that the policy trained with pure behavior cloning consistently failed. In stark contrast, our DAgger-augmented behavior cloning method successfully traversed all gates without collision. Our findings confirm that DAgger effectively mitigates the shortcomings of behavior cloning, enabling the creation of reliable and sample-efficient navigation policies for UAVs.