UC Berkeley

In recent years, quadcopters have gained remarkable popularity in a wide range of industries. Given that the utility of quadcopters has already been extensively demonstrated, this dissertation explores changes to the design and control of conventional quadcopters which either enable operation in new environments, allow for novel tasks to be performed, or reduce computational hardware requirements.

Two novel designs are presented in this dissertation. First, we explore the use of angular momentum to reduce the sensitivity of a vehicle to torque disturbances. We show both theoretically and experimentally how torque disturbance sensitivity monotonically decreases with increasing net angular momentum of the vehicle when using an appropriately designed controller, and discuss how this effect scales with vehicle size. Second, we consider the use of passive (i.e. unactuated) mechanisms to expand the types of tasks a quadcopter can perform. Specifically, we demonstrate how a vehicle with freely rotating arms can change shape mid-flight, allowing for traversal of narrow passageways, simple manipulation, and perching behaviors.

Next, two algorithms are presented which enable autonomous flight in known and unknown environments. Both algorithms are focused on improving the time required to check whether a given trajectory collides with the environment, reducing the computational power requirements of onboard computers used the fly the vehicle. The first method is used to quickly determine whether a given trajectory collides with a convex obstacle, enabling the avoidance of both static and dynamic obstacles. The second method allows for operation in previously unseen environments by using depth images from an onboard sensor to represent the environment. Rectangular pyramids are used to partition the free-space of the depth image,which enable highly efficient collision checking between trajectories and the environment.

Each proposed design and algorithm is demonstrated experimentally on custom hardware, and relevant code has been made publicly available where appropriate.