Air Force Office of Scientific Research
1 January 2008 – 30 November 2010
Co-P.I. Dr. Suman Chakravorty
Total award $450,000

This project investigates a creative and bioinspired theory of learning control which is capable of addressing the essential functionalities of a morphing Micro Air Vehicle (MAV), and which is also extensible to capabilities such as flapping and perching. The objective is to address the optimal shape control of an entire air vehicle configuration as a function of flight condition, not just simple changes such as wing sweep angle or incidence angle. The project spans theory to computation to experiment, and incorporates machine learning concepts integrated with model reference adaptive control. It uses nonlinear synthesis and simulation models of appropriate fidelity validated and verified with a hardware testbed, and culminates in a flight test demonstration.

The Defense Advanced Research Projects Agency (DARPA) defines a morphing air vehicle as a platform that is able to change its state substantially (on the order of 50%) to adapt to changing mission environments, thereby providing a superior system capability that is not possible without reconfiguration. In the context of intelligent systems, three essential functionalities of a practical morphing air vehicle are:

1. When to reconfigure
2. How to reconfigure
3. Learning to reconfigure

When to reconfigure is a major issue, as the ability for a given air vehicle to successfully perform multiple missions can directly be attributed to shape, at least if aerodynamic performance is the primary consideration. Each task or mission has an ideal or optimal vehicle shape, e.g. configuration. However, this optimality criteria may not be known over the entire flight envelope in actual practice, and the mission may be modified or completely changed during operation. How to reconfigure is a problem of sensing, actuation, and control. It is important and challenging since large shape changes produce time-varying vehicle properties, and especially, time-varying moments and products of inertia. The controller must therefore be sufficiently robust to handle these potentially wide variations. Learning to reconfigure is perhaps the most challenging of the three functionalities, and the one which has received the least attention. Even if optimal shapes are known, the actuation scheme(s) to produce them may be only poorly understood, or not understood at all; life long learning for reconfiguration strategies provide a robust evolutionary response to changing needs and missions. This permits the vehicle to be more survivable, and multi-role.

Our approach combines Machine Learning and Adaptive Dynamic Inversion Control, and is called Adaptive-Reinforcement Learning Control (A-RLC). A-RLC is a control architecture and methodology for systems with a high degree of reconfigurability, such as changing shape during flight, flapping, perching, or morphing. The key difference between our approach and the very few existing approaches to morphing control lies with how learning is used. Morphing research reported in the current literature focuses on structures and actuation of at most three degrees of morphing freedom. For a morphing MAV, even if an optimal control law is known, the actuation scheme(s) to produce this capability may be only poorly understood, or not understood at all. A-RLC is capable of addressing the optimal shape control of an entire air vehicle configuration as a function of flight condition, not just simple changes such as wing sweep angle or incidence angle. A-RLC uses Structured Adaptive Model Inversion as the trajectory tracking controller for handling time-varying time varying inertias, large variations in aerodynamic and structural properties, parametric uncertainties, and disturbances. A-RLC uses Reinforcement Learning for learning the optimality relations between the operating conditions and the desired shape, over the lifespan of the vehicle. The Reinforcement Learning module has no prior knowledge of the relationship between commands and the dimensions of the vehicle, and it does not know the relationship between the flight conditions, costs and the optimal shapes. However, the Reinforcement Learning module does know the set of all possible inputs that can be applied. From complete ignorance of the system dynamics and actuation, A-RLC is capable of learning the optimal control policy (commands) which produce the optimal shape as a function of flight condition, while maintaining accurate flight path tracking. In addition, the Reinforcement Learning module of A-RLC can function in real-time, which results in robustness with respect to model errors and environmental disturbances during system operation. Our preliminary research has demonstrated that A-RLC works well for several nonlinear, time-varying, aerodynamically effected models. Key issues we will investigate are learning and control of the morphing, aeroelastic effects, hysteretic effects, and structural effects of the high fidelity, biologically inspired models developed in this research program.

Working with me on this program are Graduate Research Assistants:

• Amanda Lampton, Ph.D. student
• Anshu Narang, Ph.D student
• Adam Niksch, M.S. student
• Kenton Kirkpatrick, M.S. student
• Monika Marwaha, M.S. student

and Undergraduate Research Assistants:

• Brian Eisenbeis
• Clark Moody
• Claire Hazelbaker