Review of modeling and control in UAV autonomous maneuvering flight Renshan Zhang
Jiyang Zhang
Huangchao Yu
Department of Unmanned Systems National University of Defense Technology Changsha, 410073, China
[email protected]
Department of Unmanned Systems National University of Defense Technology Changsha, 410073, China
[email protected]
Department of Unmanned Systems National University of Defense Technology Changsha, 410073, China
[email protected]
Abstract - The maneuver flight of unmanned aerial vehicles (UAV) is a challenging problem. which has great practical value and academic significance. This paper provides a comprehensive survey of the state-of-the-art of modeling and control of UAV maneuvering flight and exhibits the current status. The topics include maneuver actions definition and classification, modeling and control technology of maneuver UAV. Finally, some new trends for the realization of UAV maneuver flight were summarized. Index Terms - unmanned aerial vehicles ; autonomous maneuvering flight; maneuvering action; maneuvering flight control
I. INTRODUCTION Currently, unmanned aerial vehicles (UAVs) are widely used in aerial photography, mapping, intelligence, surveillance and reconnaissance [1], as well as aerial relays, ground strikes, jamming, and electronic countermeasures. However, most of the current UAV is used in stable flight mission, which only requires the UAV flying within limited rolling and pitching attitude. In a lot of cases, UAVs may be required to perform tasks beyond the regular attitude. With the expansion of the application and competition in the military field, maneuvering flight has become a key technique in the future development of UAV. Future UAVs, for example, will fly in complex environments such as urban buildings and forests [2,3]. The new UAV operation modes such as unmanned combat air vehicle (UCAV) aerial combat, man/unmanned joint operations, swarm are also being explored continuously [4-6], all of which put forward high requirements for UAV maneuvering flight. Maneuvering is an essential function of modern advanced aircraft, which enables UAVs to track and target fast-moving targets [7,8] in complex environments or change the flight path or height to avoid the attack, which can greatly enhance the mission capability and survivability probability of UAV. Maneuvering flight is a difficult problem, involving aircraft design, flight mechanics, flight control, intelligent decision and many other fields. In general, the implementation of maneuvering flight is divided into decision layer, planning layer and control layer, which involving path planning, demonstration learning, strategy search, feedback control and other algorithms. The maneuver action design, aircraft modeling and controller implementation are the hot topics. Traditional studies on UAVs focused on the realization of smooth flight, and more and more efforts are made on UAV
maneuver flights in recent years. Existing studies include flight simulation, actual flight tests, and implementation of control algorithms for various aircraft types. However, the definitions and descriptions of maneuver actions are not clear enough, and there is no representative framework for the realization of autonomous maneuvering flight control projects. This paper provides a comprehensive survey of the UAV autonomous maneuvering flight modeling and control. First, the status of studies on the maneuver flight of small UAVs as well as large fighters was summarized, and then the modeling of the UAV maneuvering flight was briefly introduced. Second, the control method of UAV maneuver flight is combed. Finally, the effect of deep-learning and fast embedded computing is briefly described in solving UAV maneuver flight issues. II. TYPICAL RESEARCH AND APPLICATION The autonomous maneuvering of UAVs as a challenging problem has attracted a lot of attentions. Some advance has been made in the aerobatics of small UAVs, research on UCAV and its maneuvering flights has also made some progress. A. Small UAV aerobatic flight In recent years, due to the improvement of sensor precision, airborne computing capability and control algorithm, the flight capability of small UAV has made some progress.
Fig 1. Small UAV Aerobatic Flight [11,15,16,23]
Table 1 lists the research of some organizations and gives some brief comments of these research. The target models of these institutes include fixed-wing, quadrotor, and helicopters.
TABLE 1. PARTIAL RESEARCH PROGRESS Research Method Major progress unit Trajectory learning Automatic flip, Helicopter Stanford University [9-11], roll, cycle and Reinforcement hurricane in place learning, DDP [12,13] ETH Zurich Navigation layer: Hover, nose circle, tail circle, H control, circles with Control layer: gain constant heading, scheduling [14] etc. McGill Accurate modeling Hovering, KnifeFixed University and traditional edge simulation wing control. [15,16] flight Korea Nonlinear pathSlow roll, knifeAerospace following guidance, edge, split-S University Specific force based maneuver on control [17,18] commanded path University Model predictive Roll, loop and of control [19-21] Immelmann Stellenbosch maneuver University Trajectory learning Hovering, of Oslo [22] Perching Adaptive fast openMulti-flips Quadrotor ETH Zurich loop, iterative learning [23-25] Type
From these studies, it can be concluded that autonomous maneuvering of small UAVs can be achieved by different planning and control algorithms. Among them, the helicopter has strong maneuverability and can achieve relatively complex maneuvers. For the unmanned fighter, the model is more complex and the realization of its autonomous maneuvering flight is also more difficult. B. UCAV maneuvering flight and its architecture In the 1990s, The United States proposed unmanned fighter aircraft in the military equipment development plan firstly, which aroused great concern among the military circles of various countries and raised the worldwide development of the UCAV. For example, Boeing converted the decommissioned F16 fighter into UAV and successfully conducted multiple flight tests. The U.S. military developed the X47-B unmanned combat aircraft and successfully completed a series of ground and shipborne tests. Several European countries have jointly developed the "neuron" unmanned combat aircraft and have completed preliminary test flights [26]. The Chinese Air Force converted the retired J-6 to an unmanned attack aircraft, which can attack the enemy's ground fixed targets. For manned aircraft, the large maneuver is based on the aircraft's own maneuvering characteristics, and also depends on the aviator’s skills and experience. With the development of UAV technology, unmanned fighters will replace manned combat aircraft in the future battlefield. Besides, the design of the unmanned fighter itself can refer to that of manned fighter aircraft, such as aerodynamic layout design and power system design. At the same time, since there is no need to consider the physical limits of the aviator, the unmanned fighters can perform maneuvers much more violently than the manned fighter. In this aspect, the unmanned fighter has a much greater potential for maneuvering than manned fighters [27]. Because of its complexity, maneuvering flight is different from the ordinary flight. The single UAV maneuver flight is the
basis for solving the maneuver flight problem of UAV. Here we simulate the pilot's aerial combat and summarizes the system framework of maneuvering technology. The behavior modeling problem of dog-fight divides dog-fight behavior models into battlefield perception and behavior generation. The perception of the battlefield is the perception of the aircraft sensors and the pilot on the battlefield. Behavior generation is used to generate fighting behavior. The behavior generation part can be divided into three functional layers: decision-making layer, tactical planning layer, and control layer [28]. The realization of UAV autonomous maneuver flight can also be divided into similar decision-making layer, planning layer and control layer.
Fig 2. Autonomous maneuvering flight architecture
Based on the mission requirements and the perceived battlefield situation, the decision-making layer selects the most suitable maneuver to form the maneuvering action chain from the current maneuver library according to a certain decision algorithm [29]. Ordinary flights missions are generally planned by designing waypoints and routes. Compared to ordinary flights, maneuvering flights require real-time attitude information of the aircraft. The planning level selects maneuvers according to the decision-making level, and refers to the current status of the aircraft in real time, generates corresponding control instructions so that the aircraft can fly according to the planned process which [30]. The control layer realizes the direct attitude control of the aircraft, namely the maneuvering command tracker. Based on the control instructions provided by the planning layer and the current flight status, the control layer generates a control surface for controlling the aircraft [31]. Compared with ordinary flight, maneuvering flight requires larger and quicker changes in attitude, and the requirements of attitude control are higher. Generally, the quaternion is used instead of Euler angles for flight calculation. III. MANEUVERING ACTIONS ANALYSIS AND MODELING The maneuvering actions are generally based on the pilot's flight experience. After many years of development, many maneuvering actions have been formed. A. Maneuver actions classification
Maneuvering flight is defined as the flight with the rapid change of flight status (speed, position, attitude, etc.) over time. In the actual air battle, the pilot drives the aircraft using maneuvering actions to change the current position and attitude of the aircraft, thereby to occupy a favorable offensive position or to evade the opponent's attack. The first maneuvered flight took place during World War I. Because of the weaker engine power of the aircraft in that year, the maneuvering flight usually stayed only in the horizontal plane. During the World War II, due to the advancement of engine technology, the maneuver flight in the three-dimensional space came into being. The maneuver library is generally based on the maneuver movements commonly used in manned air combat [32]. It simulates the pilot's maneuverability and stores a series of typical UAV maneuvering actions. According to the maneuvering process and characteristics, the change of control quantity with respect to dynamic parameters is given. Maneuvering flight movements within the three-dimensional space usually as the combination of a two-dimensional space maneuver, namely in the vertical surface of motor, including jump mobile, subduction and somersault motor, level plus (minus) speed motor, etc. As well as the maneuver in the horizontal plane which includes rapid rolling maneuver, small radius turning maneuver, etc. Through these basic flight movements, the complexly integrated maneuver flight can be realized, such as barrel roll, split S maneuver and so on [33]. But it should be noted that the composite maneuver flight vehicle presents a longitudinal and lateral aerodynamic characteristic of mutual coupling, which is not existing in two-dimensional plane maneuvering flight, so its control is difficult [34].
Fig 3. Maneuver actions [35]
There is also a special type of maneuvering behavior, which is over stall maneuvering, also called super maneuvering, which is one of the typical characteristics of a new generation of fighter jets. The common super maneuvering actions include Cobra maneuvering and Herbst maneuvering. For the super maneuvering, the attacking angle of the aircraft is usually very large. The dynamic characteristics exhibit a highly nonlinearity, and the range of the aerodynamic parameters varies with time. In the future, air warfare still be the coexistence of short-range air warfare and long-range air warfare. Therefore, the capability of super-mobile fighting cannot be ignored [36]. Supermaneuvering based on vector thrust is also a research hotspot in recent years. B. Maneuvering action model Maneuvering flight includes flight trajectories and maneuvering instructions, mostly based on mathematical formula and teaching methods. Due to the complexity of some maneuvers, it’s difficult to describe in the mathematical formula; the teaching method can also be called apprenticeship learning [37]. It refers to the process of the learner imitating the
behavior of an expert or controlling a strategy. It has been successfully used in robot navigation [38], UAVs, and unmanned vehicle control [39]. Mathematical formula According to the results of test flight completed by experienced pilots, an approximate curve of the maneuver is obtained, derived the trajectory model of the maneuver [40]. At the same time, the pilot's manipulation instructions are recorded to obtain the maneuver command generator. Discrete data sequence Data sequence: Combining multiple corrected presentation trajectories and their corresponding noise estimates, the desired reference trajectory was learned. Where z is the state vector of the target trajectory data sequence, y is the state vector of the demo trajectory data sequence, and is the time index of the target trajectory.
Fig 4. Trajectory learning algorithm [11]
Neural network fitting: The description of common maneuvering actions can be described by position sequences and random disturbances, which can be achieved by iterative learning control method. In the realization of the manned vehicle, the pilot mainly realizes the realization of the aircraft according to the short-term attitude perception. Considering that the tactical maneuvers have a shorter time of action, it is more appropriate to describe the maneuver movement using the time series of attitude. In reality, the maneuver data cannot be completely identical. The description of this similarity can be described by adding a random disturbance term, and the nonlinear time series can be constructed using a feedback timing network. Attitude sequences can be built using a sequential network. yk 1 f yk , yk 1 , ... , yk n (1)
Fig 5. The sequential network
Where y k is the output of the system at time k , z 1 is the delay link, 、 、 are roll, pitch, and yaw angles, respectively. Neural networks and deep learning methods can also be used to generate hover, landing and other command. This network directly outputs controller instructions, which is convenient to use but the function is limited and the antiinterference ability is poor [41].
Data collection
In actual flight, GPS, accelerometer, magnetometer, and other sensors can be used to record the position, attitude angle, and control data of the aircraft. However, maneuvering flight has high requirements on the accuracy and recording frequency of data. Because of the limitation of onboard computing power and sensor accuracy in actual flight, the acquired data is difficult to meet the requirements. Flight simulation technology is also a solution to the problem of mobile flight. The UAV's automatic flight simulation technology is economical and practical, which is helpful for the research and exploration of UAV design solutions. Common flight simulation software is FlightGear, XPlane, PhoenixRC, etc. Among them, data in X-Plane can be recorded at a speed of up to 99Hz. The recorded data includes aircraft position, attitude, acceleration, control amount, flight force and moment, wind Field velocity, etc. X-Plane can well record external inputs and internal states, and is a good source of data for modelling and controller design. IV. UAV MANEUVERING FLIGHT CONTROL SYSTEM Maneuvering flight is different from the conventional flight. Under high angles of attack and sideslip angles, the aerodynamic characteristics of the aircraft exhibit strong nonlinearity and coupling, resulting in the difficulty of the modelling and control of UAVs [42]. The aircraft maneuvering modelling and flight control theory will be summarized in the following sections. A. The main problem of aircraft modeling Modeling of an aircraft is the basis of flight control. However, the traditional model which attributed to the force and moment coefficients is not accurate. For example, the lift coefficient in the aerodynamic description should be a nonlinear function or a more complex differential equation that is difficult to accurately identify. C C CL =CL ( , q ) CL0 L + L q q (2) 2CL 2CL 2 1 2C L 2 ( ) q q 2 2 q q 2 Where C L is the lift coefficient, is the attack angle, and q is the pitch rate. General aircraft models didn’t consider changes in centroids, aircraft deformation and aerodynamic disturbances, the anti-interference ability is poor and it is difficult to accurately model. Therefore, the realization of accurate modeling is a challenge. There are several solutions to this problem that is worth exploring. At first, data driven modeling
can identify and represent the nonlinear dynamic character from the data directly and thus avoid the problem of approximation from the mathematical formula itself. The flexible and simple model expanded in the selected state space and time point can approximate and trace the complex dynamics in any accuracy. In these methods, flight data can be collected for online compensation to reduce the unmodeled factors. On this basis, we can use the neural network to identify the aircraft model. The neural network has strong learning and adaptive ability, and the ability to model from data, and can approximate any nonlinear function. It is very suitable for identifying complex models [43]. B. Flight control theory Most of traditional flight control methods use the classical frequency domain method or root trajectory method based on transfer function model. This control method uses modeling, identification, controller design, parameter tuning and other steps to complete the realization and synthesis of the controller. Its parameter setting is strongly dependent on experience, which brings a big risk to the flight test. Due to the limitations of classical control methods, modern control theory methods based on state variable model design, such as optimal control techniques, have gained some development in the design of flight control systems [44]. However, with the improvement of flight mission standard, the design of control system based on linear system model cannot guarantee the flight performance of aircraft with large Angle of attack. The aerodynamics of the aircraft exhibit strong nonlinearity and unsteadiness at this time, the aircraft movement is strongly coupled and the traditional small-disturbance linearization processing technology is no longer applicable. Therefore, a variety of nonlinear control methods have been developed. There are feedback linearization (differential geometry method and inverse system method), sliding mode variable structure control, nonlinear H optimization, backstepping control [45], adaptive control, neural network control, iterative learning control and so on. The advantages and disadvantages of these control methods are listed in Table 2 below. In these methods, the dynamic inversion can adapt well to the changes of the aircraft model and can meet the unconventional control requirements such as UAV maneuver flight [53]. After combining with learning and adaptive methods, the data modification model and on-line self-learning are used to solve unmodeled factors and random disturbances. Its combination with neural networks is a hot topic in the field of flight control research and has certain engineering application prospects. For example, this method has passed theoretical verification, spacecraft verification, X-35 verification, and finally applied to the F-35 fighter [54].
TABLE 2. COMPARISION OF NONLINEAR CONTROL METHODS Control methods
Advantages
Disadvantages
H
Good robustness and can reach the expected performance index[46]
For the structural uncertainty, there is often a conservative design
Dynamic inverse
Simple and easy for engineering applications. It can adapt well to the changes of the aircraft model and can meet unconventional control requirements such as UAV maneuver flight[47]
High requirements for model accuracy and sensor accuracy, and precise models are needed
Backstepping control
Convergence is better, and the unknown disturbances with uncertainties can be effectively processed[48]
The problem of robustness and actuator saturation still needs further study
Sliding mode variable structure control
Responds quickly, and the precision of the established mathematical model is not high, and it is insensitive to external uncertain perturbation reactions[49]
Necessary to further improve the ability to solve the chattering phenomenon
Adaptive control
Can modify own characteristics to adapt to changes in objects, disturbances or the environment[50]
The amount of calculation and real-time performance of the system solution need to be solved
Neural network control
Self-organization selflearning ability, strong fault tolerance and robustness[51]
Stability needs improvement; learning rate is slow and it is difficult to meet realtime control
Iterative learning control
Requires less prior knowledge of the controlled system, can approximate the desired trajectory with arbitrary precision, and the controller can learn to achieve online optimization[52]
The algorithm is sensitive to the initial value deviation; the controller lacks sufficient generalization ability, etc.
control
V. SUMMARY AND PROSPECT With the increment of the mission requirements of the aircraft and the complexity of the design of the aircraft itself, higher requirements are improved on the entire control system. In recent years, the research on hardware and software of unmanned systems has developed rapidly. The development of flight control has gone through traditional control, modern control, and nonlinear control, and has achieved many breakthroughs, but it still cannot meet the task requirements of complex systems. The improvement of sensors, intelligent control, and development of neural network chips have brought new possibilities. Control of autonomous maneuvering flights and complex systems is expected to achieve breakthroughs in these areas. Sensors and Test technology: For example, MEMS inertial
sensors such as MEMS gyroscopes and MEMS accelerometers have rapidly developed in recent years. It has the characteristics of high integration, low power consumption, and low cost. The accuracy of MEMS gyroscopes is just 1°/s when they first appear. At this stage, the MEMS gyro drift can reach within 1°/h. The development of miniature attitude measurement systems has also made new progress. With the improvement of sensor accuracy and the increasing miniaturization, a large amount of manual flight experience can be accurately recorded. UAV design, modeling, and flight testing can be achieved through data drive [55]. Big data and deep learning: the deep neural network can obtain deep-level feature representations, eliminate the complexity of the manually selected features and the dimensional disaster problem of high-dimensional data. For deep high-dimensional data control systems, the introduction of deep learning has a certain significance. Such as deep fuzzy control network, the combination of the adaptive dynamic programming (ADP) and deep learning, the use of deep ReLU network and RNN to identify dynamic models of complex unmanned systems [56]. The deep network training usually requires the support of a large amount of sample data, and the mutual compensation of big data and deep learning brings new possibilities for maneuvering flight. The development of neural network chips: traditional airborne equipment has some problems such as poor computing power, excessive power consumption, limited volume and weight, which seriously restricts the performance of UAVs. In recent years, neuro-morphological chips have received a lot of attention, such as the Neurogrid of Stanford University, TrueNorth of IBM, Zeroth of Qualcomm, and Cambrian chips of Chinese Academy of Sciences [57,58]. Compared to neural networks that are calculated and implemented using generalpurpose processors and software, neural network chips can achieve very high real-time budgets with extremely low power consumption and can meet real-time information processing needs. After years of development, UAV autonomous maneuvering has made great breakthroughs. This article mainly introduces the planning and control layers of maneuvering flights. With the development of artificial intelligence and the enhancement of airborne computing capability, the autonomous flying capability of UAVs in the future will surely be greatly improved, and it is expected to be successfully applied to the battlefield. REFERENCES [1] Y. FAN, "Autonomous and intelligent control of the unmanned aerial vehicle," SCIENTIA SINICA Technologica, vol. 47, no. 3, pp. 221-229, 2017. [2] A. Bry, C. Richter, A. Bachrach, and N. Roy, "Aggressive flight of fixedwing and quadrotor aircraft in dense indoor environments," The International Journal of Robotics Research, vol. 34, no. 7, pp. 969-1002, 2015. [3] A. A. Paranjape, K. C. Meier, X. Shi, S.-J. Chung, and S. Hutchinson, "Motion primitives and 3D path planning for fast flight through a forest," The International Journal of Robotics Research, vol. 34, no. 3, pp. 357-377, 2015. [4] F. Li and Z. Kun, "Study on UCAV air-to-ground Attack's key technologies," in Control and Decision Conference (CCDC), 2011 Chinese, 2011, pp. 4077-4081: IEEE. [5] R. Osterhuber, "FCS Requirements for Combat Aircraft—Lessons Learned for Future Designs," in Proceedings of the NATO RTO AVT-189 Specialists
Meeting, 2011. [6] A. Brown, J. Dillon, G. Craig, and R. Erdos, "Flight Manoeuvre and Spin Characteristics of the Harvard 4: Application to Human Factors Flight Research," in AIAA Atmospheric Flight Mechanics Conference and Exhibit, 2004, p. 4815. [7] E. Frazzoli, "Maneuver-based motion planning and coordination for single and multiple UAVs," in 1st UAV Conference, 2002, p. 3472. [8] S. Park, "Avionics and control system development for mid-air rendezvous of two unmanned aerial vehicles," Massachusetts Institute of Technology, 2004. [9] Tang, J., et al. Parameterized maneuver learning for autonomous helicopter flight. in IEEE International Conference on Robotics and Automation. 2010. [10] Coates, A., P. Abbeel and A.Y. Ng. Learning for control from multiple demonstrations. in International Conference on Machine Learning. 2008. [11] Abbeel, P., A. Coates and A.Y. Ng, Autonomous Helicopter Aerobatics through Apprenticeship Learning. International Journal of Robotics Research, 2010. 29(13): p. 1608-1639. [12] Abbeel, P., et al. An application of reinforcement learning to aerobatic helicopter flight. in International Conference on Neural Information Processing Systems. 2006. [13] Ng, A.Y., et al., Autonomous Inverted Helicopter Flight via Reinforcement Learning. 2006. 21(4): p. 363-372. [14] M. B. Gerig, "Modeling, guidance, and control of aerobatic maneuvers of an autonomous helicopter," ETH Zurich, 2008. [15] Bulka, E. and M. Nahon. Autonomous control of agile fixed-wing UAVs performing aerobatic maneuvers. in International Conference on Unmanned Aircraft Systems. 2017. [16] W. Khan, "Dynamics Modeling of Agile Fixed-Wing Unmanned Aerial Vehicles," McGill University Libraries, 2016. [17] Park, S. Autonomous Aerobatic Flight by Three-Dimensional PathFollowing with Relaxed Roll Constraint. in AIAA Guidance, Navigation, and Control Conference. 2013. [18] Park, S., Autonomous aerobatics on commanded path. Aerospace Science & Technology, 2012. 22(1): p. 64-74. [19] Hough, W.J., Autonomous aerobatic flight of a fixed wing unmanned aerial vehicle. Stellenbosch University of Stellenbosch, 2007. [20] Peddle, I.K., Acceleration based manoeuvre flight control system for unmanned aerial vehicles. Stellenbosch Stellenbosch University, 2008. [21] Peddle, I.K. and T. Jones, Acceleration-based 3D Flight Control for UAVs: Strategy and Longitudinal Design. 2011: InTech. [22] Sevaldson, M.C., Trajectory Learning for Highly Aerobatic Unmanned Aerial Vehicle. 2012. [23] Lupashin, S. and R. D Andrea, Adaptive fast open-loop maneuvers for quadrocopters. Autonomous Robots, 2012. 33(1-2): p. 89-102. [24] Lupashin, S. and R. D'Andrea, Adaptive Open-Loop Aerobatic Maneuvers for Quadrocopters. IFAC Proceedings Volumes, 2011. 44(1): p. 2600-2606. [25] Ritz, R. and R. D'Andrea. An on-board learning scheme for open-loop quadrocopter maneuvers using inertial sensors and control inputs from an external pilot. in IEEE International Conference on Robotics and Automation. 2014. [26] Simon, P., Military Robotics: Latest Trends and Spatial Grasp Solutions. International Journal of Advanced Research in Artificial Intelligence, 2015. 4(4). [27] Wyatt, E. The DARPA/Air Force Unmanned Combat Air Vehicle (UCAV) Program. in AIAA International Air and Space Symposium and Exposition: The Next 100 Years. 2013. [28] Kaneshige, J., K. Krishnakumar and F. Shung, Tactical Maneuvering Using Immunized Sequence Selection. Aiaa Journal, 2006. [29] Zhou, S.Y., et al., Overview of Autonomous Air Combat Maneuver Decision. Aeronautical Computing Technique, 2012. [30] Wang, J. and Z.H. Gao, The Automatic Flight Simulation of Waypoint Flight. Flight Dynamics, 2008. [31] Wang, J., Z.H. Gao and C.Y. Shang, Design of maneuvering command tracker in flight simulation. Electronics Optics & Control, 2008. [32] Austin, F., et al., Automated maneuvering decisions for air-to-air combat. Aiaa Journal, 2013. [33] https://en.wikipedia.org/wiki/Basic_fighter_maneuvers [34] Chen, Y.R. and J.P. Yuan, Research on Matching Suitability of Geomagnetism Map Based on Fractal Dimension. Flight Dynamics, 2009. 27(6): p. 76-79. [35] http://bbs.tiexue.net/post2_5713654_1.html [36] Gavrilets, V., et al., Aggressive Maneuvering of Small Autonomous Helicopters: A Human-Centered Approach. International Journal of Robotics
Research, 2001. 20(10): p. 795-807. [37] Coates, A., P. Abbeel and A.Y. Ng, Apprenticeship learning for helicopter control. Communications of the Acm, 2010. 52(7): p. 97-105. [38] Abbeel, P. Apprenticeship learning and reinforcement learning with application to robotic control. 2008. [39] Abbeel, P. and A.Y. Ng. Apprenticeship learning via inverse reinforcement learning. in International Conference on Machine Learning. 2004. [40] Yang, Q.W. and Z.Y. Zhan, Design of Command Generator for UCAV. Journal of Air Force Engineering University, 2010. [41]. Wyeth, G.F., G.D. Buskey and J. Roberts. Flight control using an artificial neural network. in ACRA 2000. 2013. [42] R.-g. Zhu, C.-s. Jiang, Q.-y. Zou, and S.-l. Cai, "Study on dynamic inversion control and simulation of super-maneuverable flight of the new generation fighter," ACTA AERONAUTICA ET ASTRONAUTICA SINICASERIES A AND B-, vol. 24, no. 3, pp. 242-245, 2003. [43] R. K. Chauhan and S. Singh, "Application of neural networks based method for estimation of aerodynamic derivatives," in International Conference on Cloud Computing, Data Science & Engineering - Confluence, 2017, pp. 5864. [44] Ross, I.M. and M. Karpenko, A review of pseudospectral optimal control: From theory to flight. Annual Reviews in Control, 2012. 36(2): p. 182-197. [45] Wang, M.X., L.I. Ming and Z.J. Zhang, Developing Status of Control Law Design Methods for Flight. Flight Dynamics, 2007. 25(2): p. 1-4. [46] H. Liu, W. Zhao, Z. Zuo, and Y. Zhong, "Robust control for quadrotors with multiple time-varying uncertainties and delays," IEEE Transactions on Industrial Electronics, vol. 64, no. 2, pp. 1303-1312, 2017. [47] Wacker R, Munday S, Merkle S. X-38 application of dynamic inversion flight control[J]. 2001. [48] L. Cao, X. Hu, S. Zhang, and Y. Liu, "Robust Flight Control Design Using Sensor-Based Backstepping Control for Unmanned Aerial Vehicles," Journal of Aerospace Engineering, vol. 30, no. 6, p. 04017068, 2017. [49] Z. Su, H. Wang, N. Li, Y. Yu, and J. Wu, "Exact docking flight controller for autonomous aerial refueling with back-stepping based high order sliding mode," Mechanical Systems and Signal Processing, vol. 101, pp. 338-360, 2018. [50] C. Cao and N. Hovakimyan, "L1 adaptive output feedback controller for systems with time-varying unknown parameters and bounded disturbances," in American Control Conference, 2007. ACC'07, 2007, pp. 486-491: IEEE. [51] O. Dadian, S. Bhandari, and A. Raheja, "A recurrent neural network for nonlinear control of a fixed-wing uav," in American Control Conference (ACC), 2016, 2016, pp. 1341-1346: IEEE. [52] M. Zhaowei, H. Tianjiang, S. Lincheng, K. Weiwei, Z. Boxin, and Y. Kaidi, "An iterative learning controller for quadrotor UAV path following at a constant altitude," in Control Conference (CCC), 2015 34th Chinese, 2015, pp. 44064411: IEEE. [53] Burken, J.J., et al., Adaptive Control Using Neural Network Augmentation for a Modified F-15 Aircraft. 2006. 1-6. [54] G. P. Walker, J. W. Fuller, and S. P. Wurth, "F-35B integrated flightpropulsion control development," in 2013 International Powered Lift Conference, 2013, p. 4243. [55] E. Denti, R. Galatolo, and F. Schettini, "An AHRS based on a Kalman filter for the integration of inertial magnetometric and GPS data," in 27th International Congress of the Aeronautical Sciences, 2010. [56] D. Yan-Jie, L. Yi-Sheng, Z. Jie, Z. Xue-Liang, and W. Fei-Yue, "Deep learning for control: the state of the art and prospects," Acta Automatica Sinica, vol. 42, no. 5, pp. 643-654, 2016. [57] S. Furber and S. Temple, "Neural systems engineering," Journal of the Royal Society interface, vol. 4, no. 13, pp. 193-206, 2007. [58] F. Akopyan et al., "Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip," IEEE Transactions on ComputerAided Design of Integrated Circuits and Systems, vol. 34, no. 10, pp. 1537-1557, 2015.