Kinodynamic planning is motion planning in state space and aims to satisfy kinematic and dynamic constraints. To reduce its computational cost, a popular approach is to use sampling based methods such as RRT with off-line machine learning for estimating the steering cost and inputs. However, scalability and robustness are still open challenges in these type of Learning-RRT algorithms. We propose the use of generative adversarial networks (GAN) for learning of the steering cost and inputs. Furthermore, a novel data generation method is introduced, which is easy to learn and, in terms of parameter count, scales linearly to higher degrees of freedom. In our experiments, we show that the GAN has excellent generalisation capabilities, resulting in a considerable improvement in performance compared to the state-of-the-art. Consequently, we show that our method can scale to a planar arm and is robust to data dimensionality.

On-board stabilization of quadrotors is often done using an Inertial Measurement Unit (IMU), aided by additional sensors to combat the IMU drift. For example, GPS readings can aid when flying outdoors, or when flying in GPS denied environments, such as indoors, visual information from one or more camera modules can be used.
A single downwards facing camera however cannot determine the absolute height of the quadrotor, leaving the results from the Optical Flow (OF) up to scale. To estimate the velocity of the quadrotor an additional range sensor, such as an Ultrasonic Sensor (US), is used to solve this scaling problem.
These solutions are difficult to scale down to micro quadrotors as the platform becomes too small to fit and lift additional sensors. Therefore stabilizing a quadrotor with a single camera and IMU only would pave the way for the development of even smaller quadrotors. This master thesis presents an adaptive control strategy to stabilize a micro quadrotor in all
three axes using only an IMU and a monocular camera. This is achieved by extending the stability based approach for a single, vertical, axis by De Croon in Distance estimation with efference copies and optical flow maneuvers: a stability-based strategy[1]. This stability based method ncreases the control gain in the visual feedback loop until the quadrotor detects it is oscillating by detecting that the covariance of the given thrust inputs and the measured divergence passes a threshold. Next the height can be estimated using the predetermined relationship between gain and height at which these self-induced oscillations occur and proper gains can be set for the estimated height.
An analysis is done in simulation to present proof of concept of the stabilization method in three axis and to determine the effects of scaling and the effects of varying effective Frames per Second (FPS) caused by computations. It was shown that the adaptive gain strategy can stabilize the simulated quadrotor and prevent it from drifting. Furthermore, the control gains were scaled such that the effects of scaling a quadrotor could be mostly negated, though at about a tenth of the scale the simulated noise had such an influence that the scaled gains could not negate it anymore. Furthermore, the minimum effective FPS required to stabilize an ARDrone 2 was determined to be 15 FPS, and it was shown that an increase in effective FPS aids stabilizing the smaller scale quadrotors that became unstable due to the scaling effects.
Furthermore, flights on an Parrot ARDrone 2 and Parrot Bebop are performed to show the usability of this control strategy in real life. It was shown that both quadrotors could achieve stable hover without drifting at multiple heights, using various strategies.