Download Methods for Online Predictive Control of Multi-rotor Aerial Robots with Perception-driven Tasks Subject to Sensing and Actuation Constraints Book in PDF, Epub and Kindle
Drones have an increasing place in numerous applications already started to take advantage from those, in particular in the fields of photography and video making, or simply for leisure activities. Simultaneously, the picture of autonomous aerial robots widely spread as a mark of innovation, such that many civilian of industrial applications are now envisioned through this aspect. One could cite, for instance, the persistent idea of aerial home delivery of goods, exploited by many companies. Another spread use-case is the deployment of fleets of aerial robots for monitoring activities, in hard-to-access environments, such as high mountains.The aerial robotics research community is active from numerous years, and the state of the art keeps improving, being through the conception of novel, more adaptive control algorithms, or the improvements of the hardware designs, opening new ranges of possibilities.The deployment of such robots in the scope of applications in uncontrolled environments comes with a lot of challenges, in particular regarding the perception of the surroundings. Exteroceptive sensors are indeed mandatory for most of autonomous applications. Among those sensors, cameras hold a peculiar position.It is on the one hand due to the simple onboard integration with their small size and weight,and on the other hand to the design of human-made environments, which are heavily built around visual markers (signs, illuminated signals...) However, maintaining visibility over objects or phenomenon often collide with the motion requirements of the robot, or with the tasks to which it is assigned. This effect is prominent when using underactuated robots, which are the most widely spread types of aerial vehicles, partly because of their higher energy efficiency. This property implies a strong coupling between position and orientation: the robot needs to tilt to move, and corollary moves when it tilts, thus altering the sensor bearing.From this assessment, the robotics community works to produce sensorimotor algorithms, able to produce motions while accounting for perception.This thesis takes place in this context, aiming at proposing such control methods to enforce the visibility over a phenomenon of interest through the onboard sensors. Moreover, to ensure the feasibility of the generated commands, it is required to account for the various actuation limitations of the robots. Finally, this thesis devotes to propose generic formulations, thus avoiding to propose ad hoc solutions, which would be contingent to a specific problem.To tackles these aspects under a common formalism, the proposed solutions are based on optimal and predictive control policies. These are based on numerical optimization, implying the need of accurate models, and thus accounting for the system nonlinearities, which are often disregarded for simplification.The contributions of this these are the aggregation of the various concepts in a common paradigm,and the formalization of the various mathematical functions transcribing the objectives and constraints related to perception. This paradigm is used in the scope of several applications related to usual perception-driven tasks in aerial robotics, namely the tracking of dynamic phenomenon, the improvement of this tracking, or the visual-inertial localization. Finally, the proposed solutions are implemented and tested in simulations and on real aerial robots.The work conducted throughout this thesis led to various publications in international peer-reviewed conferences and journals. All the related software production from these works are published open-source for the robotics community.