The idea of integrating a controller with perception, control and planner in a single network is explored. A simulation environment was constructed using RAI, a c++ framework for benchmark and development of learning algorithms. A policy was trained which uses the depth of the scene and the state of the quadrotor and outputs rotor thrust. The policy has shown that it is capable of avoiding obstacles with a high success rate. Moreover, it was shown that the policy can adapt to camera parameters such as depth range, frame update rates.
A slung load on a quadrotor can be used for various applications as load transportation, aerial manipulation. As it is the most minimal setup for manipulating objects it can be agile and efficient compared to actuated manipulators. However, the dynamics are nonlinear and hybrid, thus it is hard to plan and evaluate trajectorties for collision. This project uses a direct method in order to optimize a collision free trajectory which uses 3D minimum distances to evaluate collision. The result is a much faster planning in collision free trajectory planning which can handle a large number of obstacles.
Nguyen(2016) showed that a 3 axis force can be generated from a tool rigidly attached to a underactuated quadrotor. This project implements the theory into a real platform.
SNUSAT-1 is a 2U Cubesat developed as part of the QB50 EU project with 50 other universities. My contribution was in the electrical system of the satellite. I also developed a PCB type magnetic torquer, which is a low power actuator used for controlling the attitude of the satellite. SNUSAT-1 and SNUSAT-1b was launched to orbit on July 2017.