Learning safe recovery trajectories with deep neural networks for unmanned aerial vehicles

Unmanned vehicles that use vision sensors for perception to aid autonomous flight are a highly popular area of research. However, these systems are often prone to failures that are often hard to model. Previous work has focused on using deep learning to detect these failures. In this work, we build on these failure detection systems and develop a pipeline that learns to identify the correct trajectory to execute that restores the vision system and the unmanned vehicle to a safe state. The key challenge with using a deep learning pipeline for this problem is the limited amount of training data available from a real world system. Ideally one requires millions of data points to sufficiently train a model from scratch. However, this is not feasible for an unmanned aerial vehicle. The dataset we operate with is limited to 400–500 points. To sufficiently learn from such a small dataset we leverage the idea of transfer learning and non linear dimensionality reduction. We deploy our pipeline on an unmanned aerial vehicle flying autonomously through outdoor clutter (in a GPS denied environment) and show that we are able to achieve long durations of safe autonomous flight.