Publications (9)6.2 Total impact
- [Show abstract] [Hide abstract]
ABSTRACT: The appearance variation cue captures the variation in texture in a single image. Its use for obstacle avoidance is based on the assumption that there is less such variation when the camera is close to an obstacle. For videos of approaching frontal obstacles, it is demonstrated that combining the cue with optic flow leads to better performance than using either cue alone. In addition, the cue is successfully used to control the 16-g flapping-wing micro air vehicle DelFly II.IEEE Transactions on Robotics 01/2012; 28(2):529-534. · 2.57 Impact Factor
Article: Polynomial inclusion functionsReliable Computing 01/2012; 16:283-307. · 0.68 Impact Factor
Conference Paper: The appearance variation cue for obstacle avoidance[Show abstract] [Hide abstract]
ABSTRACT: A visual cue is introduced that exploits the visual appearance of a single image to estimate the proximity to an obstacle. In particular, the appearance variation cue captures the variation in texture and / or color in the image, and is based on the assumption that there is less such variation when the camera is close to an obstacle. Random sampling is applied in order to evaluate the appearance variation fast enough for use in robotics. It is demonstrated that the randomly sampled appearance variation cue can be complementary to optic flow for obstacle detection; combining the two visual cues leads to better obstacle detection performance. Random sampling leads to sufficient computational efficiency for the cue's utilization in autonomous flight: a speed-up of a factor ~100 is attained, which allows the successful control of the 16-gram flapping wing MAV DelFly II.Robotics and Biomimetics (ROBIO), 2010 IEEE International Conference on; 01/2011
- Journal of Guidance Control and Dynamics - J GUID CONTROL DYNAM. 01/2010; 33(1):42-52.
- [Show abstract] [Hide abstract]
ABSTRACT: The problem of output optimization within a specified input space of neural networks (NNs) with fixed weights is discussed in this paper. The problem is (highly) nonlinear when nonlinear activation functions are used. This global optimization problem is encountered in the reinforcement learning (RL) community. Interval analysis is applied to guarantee that all solutions are found to any degree of accuracy with guaranteed bounds. The major drawbacks of interval analysis, i.e., dependency effect and high-computational load, are both present for the problem of NN output optimization. Taylor models (TMs) are introduced to reduce these drawbacks. They have excellent convergence properties for small intervals. However, the dependency effect still remains and is even made worse when evaluating large input domains. As an alternative to TMs, a different form of polynomial inclusion functions, called the polynomial set (PS) method, is introduced. This new method has the property that the bounds on the network output are tighter or at least equal to those obtained through standard interval arithmetic (IA). Experiments show that the PS method outperforms the other methods for the NN output optimization problem.IEEE Transactions on Neural Networks 05/2009; · 2.95 Impact Factor
- Navigation, Journal of the Institute of Navigation. 01/2009; 56(3):205-219.
Article: TRANSMITTER-RECEIVER SYSTEM[Show abstract] [Hide abstract]
ABSTRACT: Abstract of WO 2009095134 (A1) The invention relates to a transmitter-receiver system comprising at least three transmitters and at least a first receiver and a second receiver, wherein the receivers are connected to a computing device that is arranged to analyse signals that said receivers receive from said transmitters and to calculate length and attitude information of an imaginary baseline connecting said receivers depending on at least carrier phase information of said signals using interval analysisWO 2009095134 (A1). 01/2009;
Conference Paper: New approach for integer ambiguity resolution using interval analysis[Show abstract] [Hide abstract]
ABSTRACT: Cited By (since 1996):2, Export Date: 11 February 2014, Source: Scopus21st International Technical Meeting of the Satellite Division of the Institute of Navigation, ION GNSS 2008; 01/2008
- [Show abstract] [Hide abstract]
ABSTRACT: In the field of reconfigurable flight control (RFC) one can distinguish two types of failures. One type deals with actuator failures and the other with the more general case of structural failures. Many researchers have successfull y implemented RFC systems that can cope with actuator failures ((3), (4), (7), etc.). However, the techniques use d in these RFC systems are often not applicable to the second type of failures. Many RFC systems use some form of failure detection and isolation/identification (FDI) appr oach based on a predetermined finite set of failure cases. In case of structural failures, however, the number of possibl e failure cases is infinite and the corresponding flight dynami c characteristics are very hard to predict. In this paper a RFC system based on non-linear dy- namic inversion (NDI) (5) (1) in combination with on-line aerodynamic model identification is proposed. This type of RFC system circumvents the need for predetermined failure cases by using neural networks to keep the internal model of the NDI controller up to date. The assumption is made that the aerodynamic effects of structural failures ar e much more important compared to the changes of mass, mass distribution and the rotational inertias. This means that only identification of the total aerodynamic force and moment coefficients input-output mappings is necessary. On-line estimation of the states and total aerodynamic force and moment coefficients (using an extended Kalman filter (EKF)) provides the input-output data for the on-line identification of the aerodynamic model. The main advan- tage of using neural networks is that the internal structure of the aerodynamic model does not have to be defined. Keeping the input span of the neural networks general will ensure that all possible dependencies between input and output can be learned, resulting in an accurate aerodynamic model. A drawback of neural networks is that they suffer from the 'recency'-effect. Neural networks tend to 'forget ' knowledge which has been learned in the past if that knowledge is no longer in the recent input-output (IO) pairs presented to the network (2). The larger the input/output dimensions and range, the larger this effect. To reduce the recency effect the aerodynamic model is defined by a nine dimensional hyperbox structure. Each hyperbox contains neural networks with smaller input/output dimensions and range, which learn a small local IO-mapping of the aerodynamic model. Experiments have been performed using the non-linear model of the F-16 aircraft taken from (6). Experimental re- sults show that an accuracy of > 99% in the approximation of the aerodynamic model can be obtained throughout the flight envelope with off-line learning using the Levenberg- Marquardt algorithm. Wind tunnel test data has been used for the off-line training and subsequent validation. NDI control using the off-line learned aerodynamic model is shown to be stable and good performance is possible. The NDI is, however, sensitive to modeling errors and sensor noise. A more robust version of the used NDI control system is needed to improve performance. Further research is needed to investigate the on-line performance of the proposed RFC system.
Delft University Of Technology
Delft, South Holland, Netherlands
- Control and Simulation