Abstract
The end goal of a reactive flight control pipeline is to output control commands based on local sensor inputs. Classical state estimation and control algorithms break down this problem by first estimating the robot’s velocity and then computing a roll and pitch command based on that velocity. However, this approach is not robust in geometrically degenerate environments which do not provide enough information to accurately estimate vehicle velocity. Recent work has shown that learned end-to-end policies can unify obstacle detection and planning systems for vision-based systems. This work applies similar methods to learn an end-to-end control policy for a lidar equipped flying robot which replaces both the state estimator and controller while leaving long term planning to traditional planning algorithms. Specifically, this work demonstrates the feasibility of training such a policy using imitation learning and RNNs to map directly from lidar range measurements to robot accelerations without an explicit state estimate. The policy is fully trained on simulated data using procedurally generated environments, achieving an average of over 1.7km mean distance between collisions. Additionally, various real-world flight tests through tunnel and tunnel-like environments demonstrate that a policy learned in simulation can successfully control a real quadcopter.

Speaker Bio
Sam Zeng is an M.S. student in the Robotics Institute at Carnegie Mellon University advised by Dr. Sebastian Scherer. His current research is focused on applying machine learning to reactive control and state estimation for UAVs. Sam received his undergraduate degree from CMU in Mechanical Engineering with an additional major in Robotics in 2016.