Deep Imitation Learning for Safe Indoor Autonomous Micro Aerial Vehicle Navigation

Drones or MAVs are cyber-physical systems that have become ideal platforms for many applications in outdoor settings. Autonomous navigation in these outdoor settings has been effectively done using global positioning systems (GPS). Many human MAV pilots have demonstrated skills in controlling MAVs to maneuver in narrowed spaces indoors. However, this skill is hard to automate. Global positioning systems are unreliable indoors and in cluttered and confined environments. SLAM, being the most widely used analytical method, addresses the navigation problem by rendering a spatial map of an environment in which the Agent navigates while simultaneously localizing the Agent relative to this map. The main downside of SLAM is that rendering an entire map requires a large amount of computation, and rule-based techniques often lose their robustness in corner cases or situations that were not accounted for in developing the rules. Hence, learning directly from human demonstrations could produce improved results for complex tasks, particularly in sensor-limited systems. In this study, a policy that safely navigates a MAV through an indoor environment is learned through deep imitation learning. To effectively learn a policy that is robust to the domain or environment shifts, an ideal combination of monocular depth estimate and dense optical flow was determined to serve as state representation. Three different deep convolutional neural network architectures, namely, CNN, LSTM-RNN, and 3D CNN, were explored and developed to encode the navigation policy from expert demonstrations. The performance of these policies was then tested in a real environment. Results show that the CNN and 3D CNN policies successfully navigated the MAV around the obstacle set in the test environment while the LSTM-RNN did not. The Success Rate for CNN, LSTM-RNN, and 3D CNN were 90%,0%, and 90%, respectively.