February 12 (Thursday) John Roberts (MIT, Computer Science and Artificial Intelligence (CSAIL))
Learning Control at Intermediate Reynolds Numbers: Experiments with Flapping Flight and Perching Aircraft
Abstract:
In this talk, I'll describe some of the MIT Robot Locomotion Group's new work using machine learning to design nonlinear control systems for fluid-body interactions. The first problem I'll discuss is the optimization of the performance of a 'heaving foil' built by Jun Zhang at NYU. As of yet, we have not identified a compact and rich lumped-parameter model for this system, but I will present results demonstrating that flapping efficiency (defined as the dimensionless cost-of-transport) can be optimized on the laboratory experimental system in less than ten minutes using a model-free optimal control algorithm. To perform the optimization this quickly we made use of novel algorithmic improvements motivated by a Signal-to-Noise Ratio we developed for policy gradient algorithms. I will then briefly describe the lab's work on making a fixed-wing aircraft execute a post-stall maneuver in order to land on a perch. Experimental system identification provides a compact lumped parameter model, allowing us to apply a model-based optimal control algorithm with promising results and investigate how the controllability of the perching maneuver depends upon the type of actuation available to the plane. I will conclude by discussing how these advances may be used to make a robotic ornithopter built by our lab more capable of flying and landing like a bird.