In the context of the PortWings project (http://www.portwings.eu/), we designed a sophisticated set-up able to actuate and sense forces and torques in different spatial directions.
In particular, the tool, or end-effector, that the torques are transmitted to, is supposed to be an aerodynamic profile similar to a wing, and a 2-DOF actuation system serves to induce a flapping motion on the wing. A 6-DOF sensor placed at the joint of the wing is able to sense forces and torques in all spatial directions.
This assignment aims at developing a complete study of the system including:
-Modeling and Simulation: Developing a dynamic model for the set-up, including aspects related to its series-elastic actuation, and being able to perform dynamic simulations on the model, which will include linear aerodynamic effects.
-Control: Nontrivial control strategies can be implemented to achieve the desired task, which will deal with performing optimal flapping motions, and which can be tested in simulation and possibly on the real set-up. Given the highly complicated nature of the involved dynamics, we aim at using reinforcement learning in order to solve hard optimisation problems using data coming from the simulated dynamics. Different learning architectures can be compared, which can range across different action spaces (the system can be controlled in position, velocity and toque) and different possible feedbacks (the simulating environment would provide information on the state of the system).
As a final goal, we aim at using the learned controller in simulation as an initial guess for performing the same learning procedure on the real set-up, in which the complicated effects on the nonlinear aerodynamics and the flexibility of the wing, were not included in the model used for simulations, will play a role to fine-tune the controller. We claim in this way a faster and more reliable convergence to some minimum of the optimisation problem.