Biblio

Machine-learning and soft computation methods are often used to adapt and modify control systems for robotic, aerospace, and other electromechanical systems. Most often, those who use such methods of self-adaptation focus on issues related to efficacy of the solutions produced and efficiency of the computational methods harnessed to create them. Considered far less often are the effects self-adaptation on Verification and Validation (V{&}V) of the systems in which they are used. Simply observing that a broken robotic or aerospace system seems to have been repaired is often not enough. Since self-adaptation can severely distort the relationships among system components, many V{&}V methods can quickly become useless. This paper will focus on a method by which one can interleave machine-learning and model consistency checks to not only improve system performance, but also to identify how those improvements modify the relationship between the system and its underlying model. Armed with such knowledge, it becomes possible to update the underlying model to maintain consistency between the real and modeled systems. We will focus on a specific application of this idea to maintaining model consistency for a simulated Flapping-Wing Micro Air Vehicle that uses machine learning to compensate for wing damage incurred while in flight. We will demonstrate that our method can detect the nature of the wing damage and update the underlying vehicle model to better reflect the operation of the system after learning. The paper will conclude with a discussion of potential future applications, including generalizing the technique to other vehicles and automating the generation of model consistency-testing hypotheses.

This paper proposes a model checking method for a trajectory tracking controller for a flapping wing micro-air-vehicle (MAV) under disturbance. Due to the coupling of the continuous vehicle dynamics and the discrete guidance laws, the system is a hybrid system. Existing hybrid model checkers approximate the model by partitioning the continuous state space into invariant regions (flow pipes) through the use of reachable set computations. There are currently no efficient methods for accounting for unknown disturbances to the system. Neglecting disturbances for the trajectory tracking problem underestimates the reachable set and can fail to detect when the system would reach an unsafe condition. For linear systems, we propose the use of the H-infinity norm to augment the flow pipes and account for disturbances. We show that dynamic inversion can be coupled with our method to address the nonlinearities in the flapping-wing control system.

Biomimetic flapping wing vehicles have attracted recent interest because of their numerous potential military and civilian applications. In this paper we describe the design of a multi-agent adaptive controller for such a vehicle. This controller is responsible for estimating the vehicle pose (position and orientation) and then generating four parameters needed for split-cycle control of wing movements to correct pose errors. These parameters are produced via a subsumption architecture rule base. The control strategy is fault tolerant. Using an online learning process an agent continuously monitors the vehicle's behavior and initiates diagnostics if the behavior has degraded. This agent can then autonomously adapt the rule base if necessary. Each rule base is constructed using a combination of extrinsic and intrinsic evolution. Details on the vehicle, the multi-agent system architecture, agent task scheduling, rule base design, and vehicle control are provided.