| Unmanned helicopters hold both tremendous potential and challenges. Without risking the lives of human pilots, these vehicles exhibit agile movement and the ability to hover and hence open up a wide range of applications in the hazardous situations. Sparing human lives, however, comes at a stiff price for technology. Some of the key difficulties that arise in these challenges are: (i) There are unexplained cross-coupled responses between the control axes on the hingeless helicopters that have puzzled researchers for years. (ii) Most, if not all, navigation on the unmanned helicopters relies on Global Navigation Satellite Systems (GNSSs), which are susceptible to jamming. (iii) It is often necessary to accommodate the re-configurations of the payload or the actuators on the helicopters by repeatedly tuning an autopilot, and that requires intensive human supervision and/or system identification.;For the dynamics modeling and analysis, we present a comprehensive review on the helicopter actuation and dynamics, and contributes toward a more complete understanding on the on-axis and off-axis dynamical responses on the helicopter. We focus on a commonly used modeling technique, namely the phase-lag treatment, and employ a first-principles modeling method to justify that (i) why that phase-lag technique is inaccurate, (ii) how we can analyze the helicopter actuation and dynamics more accurately. Moreover, these dynamics modeling and analysis reveal the hard-to-measure but crucial parameters on a helicopter model that require the constant identifications, and hence convey the reasoning of seeking a model-implicit method to solve the state estimation and control problems on the unmanned helicopters.;For the state estimation, we present a robust localization method for the unmanned helicopter against the GNSS outage. This method infers position from the acceleration measurement from an inertial measurement unit (IMU). In the core of our method are techniques of the sensor error modeling and the filtering method for the sensor noise compensation. Moreover, we provide a fully automatic algorithm to tune our method. Finally, we evaluate our method on an instrumented gasoline helicopter. Experiments show that the technique enables the robust positioning of flying helicopters when no GNSS measurement is available.;The design of an autopilot for an unmanned helicopter is made difficult by its nonlinear, coupled and non-minimum phase dynamics. Here, we consider a reinforcement learning approach to transfer motion skills from human to machine, and hence to achieve autonomous flight control. By making efficient use of a series of state-and-action pairs given by a human pilot, our algorithm bootstraps a parameterized control policy and learns to hover and follow trajectories after one manual flight. One key observation our algorithm is based on is that, although it is often difficult to retrieve the human pilots’ hidden desiderata that formulate their state-feedback mechanisms in controlling the helicopters, it is possible to intercept the states of a helicopter and the actions by a human pilot and then to fit both into a model. We demonstrate the performance of our learning controller in experiments.;The results described in this dissertation shed new and important light on the technology necessary to advance the current state of the unmanned helicopters. From a comprehensive dynamics modeling that addresses perplexing cross-couplings on the unmanned helicopters, to a robust state estimation against GNSS outage and a learn-from-scarce-sample control for an unmanned helicopter, we provide a starting point for the cultivation of the next-generation unmanned helicopters that can operate with the least possible human intervention. |
