Abstract:
The teleoperation of unmanned aerial vehicles (UAVs), especially in cramped, GPS-restricted, environments, poses many challenges. The presence of obstacles in an unfamiliar environment requires reliable state estimation and active algorithms to prevent collisions.
In this dissertation, we present a collision-free indoor navigation system for a teleoperated quadrotor UAV. The platform is equipped with an on-board miniature computer and a minimal set of sensors for this task and is self-sufficient with respect to external tracking systems and computation. The platform is capable of highly accurate state-estimation, tracking of the velocity commanded by the user and collision-free navigation. The robot estimates its state in a cascade architecture. The attitude of the platform is calculated with a complementary filter and its linear velocity through a Kalman filter integration of inertial and optical flow measurements.
An RGB-D camera serves the purpose of providing visual feedback to the operator and depth measurements to build a probabilistic, robot-centric obstacle state with a bin-occupancy filter. The algorithm tracks the obstacles when they leave the field of view of the sensor by updating their positions with the estimate of the robot's motion. The avoidance part of our navigation system is based on the Model Predictive Control approach. By predicting the possible future obstacles states, the UAV filters the operator commands by altering them to prevent collisions. Experiments in obstacle-rich indoor and outdoor environments validate the efficiency of the proposed setup.
Flying robots are highly prone to damage in cases of control errors, as these most likely will cause them to fall to the ground. Therefore, the development of algorithm for UAVs entails considerable amount of time and resources. In this dissertation we present two simulation methods, i.e. software- and hardware-in-the-loop simulations, to facilitate this process. The software-in-the-loop testing was used for the development and tuning of the state estimator for our robot using both the simulated sensors and pre-recorded datasets of sensor measurements, e.g., from real robotic experiments. With hardware-in-the-loop simulations, we are able to command the robot simulated in Gazebo, a popular open source ROS-enabled physical simulator, using computational units that are embedded on our quadrotor UAVs. Hence, we can test in simulation not only the correct execution of algorithms, but also the computational feasibility directly on the robot's hardware.
Lastly, we analyze the influence of the robot's motion on the visual feedback provided to the operator. While some UAVs have the capacity to carry mechanically stabilized camera equipment, weight limits or other problems may make mechanical stabilization impractical. With a fixed camera, the video stream is often unsteady due to the multirotor's movement and can impair the operator's situation awareness. There has been significant research on how to stabilize videos using feature tracking to determine camera movement, which in turn is used to manipulate frames and stabilize the camera stream. However, we believe that this process could be greatly simplified by using data from a UAV’s on-board inertial measurement unit to stabilize the camera feed. Our results show that our algorithm successfully stabilizes the camera stream with the added benefit of requiring less computational power. We also propose a novel quadrotor design concept to decouple its orientation from the lateral motion of the quadrotor. In our design the tilt angles of the propellers with respect to the quadrotor body are being simultaneously controlled with two additional actuators by employing the parallelogram principle. After deriving the dynamic model of this design, we propose a controller for this platform based on feedback linearization. Simulation results confirm our theoretical findings, highlighting the improved motion capabilities of this novel design with respect to standard quadrotors.