Robust Vision-Based Planning for Quadrotor UAV Using Funnel Libraries

Abstract

End-to-end vision-based solutions to robot control and planning problems have been the subject of significant research in recent years as deep neural networks started to give promising results. However, due to their lack of interpretability, providing rigorous guarantees on safety and performance of vision-based approaches is a very challenging yet crucial problem, especially when the system is subject to unknown disturbances that are not present in the training environments. Our work combines ideas from model-based reachability analysis and deep reinforcement learning to design a vision-based motion planning policy that holds PAC-Bayes generalization guarantees for planning in novel environments with unknown disturbances. The method presented in this thesis exploits the model knowledge to perform the offline computation of over-approximated reachable sets (funnels) around the motion primitive trajectories. Then, we use deep reinforcement learning to synthesize a vision-based motion planning policy that utilizes the precomputed funnel library. Finally, the PAC-Bayes framework is utilized to achieve strong generalization bounds on the average cost of the policies in novel environments. We implement this framework on a computer simulation of an autonomous quadrotor UAV navigating through obstacle-dense environments and achieve a final PAC-Bayes Bound of 20.4%. We evaluate the posterior policy distribution to validate that the PAC-Bayes generalization bound holds in novel environments with unknown external disturbances.