Data-driven model-based control for novel flying machines: From highly underactuated to overactuated systems with uncertain dynamics
Open access
Author
Date
2022Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
The central theme of this thesis is to push the limits of aerial robotics via novel vehicle design and by developing methods that combine model-based control with machine learning approach to exploit the best aspects of both methods. Conventional quadrotors have become popular due to their mechanical simplicity and agility. Meanwhile, novel vehicle designs endow flying machines with new capabilities that a standard quadrotor lacks. This thesis investigates two designs: 1. a flying vehicle which has a single moving part and is yet able to hover and fully control its position, the Monospinner; 2. a tilt-arm overactuated (18 actuators) omnidirectional flying vehicle, the Omav. For the former design, this highly underactuated flying machine aims to answer a fundamental question about flying machine: what is the minimum number of moving parts for a controllable flying vehicle. The latter design is tailored for aerial physical interaction tasks, for which flying vehicles need to exchange forces with their surrounding environment. The Omav thus has a full actuation wrench envelope (that is, the ability of producing force and torque in arbitrary directions) and is inevitably mechanically complex.
These novel designs require new modeling and control approaches: the controlled system has to be robust against internal and external disturbances, these include modeling error due to mechanical complexity, aerodynamics, manufacturing imperfections, measurement noise, and environment disturbance. While first-principle model-based control methods are powerful for design and control of dynamical systems and applied on the Monospinner, tools from statistical learning theory help to capture the part of the Omav dynamics that is hard to model using first principles. This in turn helps model-based control method to perform optimally for such a mechanical complex flying machine. Challenges arise along with this novel approach such as the data collection efficiency and efficacy and the embedding of the learned model into the controller.
For the Monospinner, the core approach is to use model based control theory to find a vehicle design that is robust under disturbance. In particular, its translational and attitude dynamics are formulated as a twelve-dimensional state space system, which may be linearized to a linear time-invariant system amenable to controllability analysis, controller synthesis, and vehicle design. A mathematical analysis is given to show the vehicle is fully controllable in position after removing its yaw state, and in particular for the case of a vehicle with the shape of a planar object and an offset thrust location (with respect to its center of mass). The equilibrium of the resulting system has a large region of attraction such that it recovers after being thrown into the air like a frisbee.
The research into the Omav puts emphasis on algorithmic methods. This thesis presents an approach that combines a data-driven and a first-principle model for the system actuation and uses it to improve the controller. Particular attention is paid to avoid ambiguous solutions present in a standard inverse model. This is solved by an optimization problem only using forward models. In addition, an efficient training data collection procedure is devised using an optimization problem formulation to find an informative trajectory. We present a sampling- based method that computes an approximation of the trajectory that minimizes the predictive uncertainty of the learned dynamics model. This trajectory is then executed, collecting the data to update the learned model. Last but not least, an adaptive control strategy is proposed for aerial sliding on surfaces with discontinuous change in geometry and unknown, spatially-varying friction properties. It augments a standard impedance controller using a control parameter adjustment policy, which combines proprioceptive measurements, tactile sensing and control signals as the policy input. In particular, this policy is trained in simulation with simplified actuator dynamics and yet is capable of being transferred to the robot without any adaptation. The key to this is the preserved controller structure.
Indoor experiments of the Monospinner and the Omav have been evaluated in various conditions. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000605660Publication status
publishedExternal links
Search print copy at ETH Library
Publisher
ETH ZurichSubject
Flying robot; Learning control; Modeling; Robot design; Model-based control; Aerial Interaction; Underactuated robots; Omnidirectional flying robotOrganisational unit
03737 - Siegwart, Roland Y. / Siegwart, Roland Y.
More
Show all metadata
ETH Bibliography
yes
Altmetrics