We seek to develop vision-based autonomy for small-scale aircraft known as Micro Air Vehicles (MAVs). Development of such autonomy presents significant challenges, in no small measure because of the inherent instability of these flight vehicles. Therefore, we propose a virtual flight testbed that seeks to mitigate these challenges by facilitating the rapid development of new vision-based control algorithms that would have been, in its absence, substantially more difficult to transition to successful flight testing. The proposed virtual testbed is a precursor to a more complex Hardware-In-the-Loop (HILS) facility currently being constructed at the University of Florida. These systems allow us to experiment with vision-based algorithms in controlled laboratory settings, thereby minimizing loss-of-vehicle risks associated with actual flight testing. In this paper, we first discuss our testbed system, both virtual and real. Second, we present our vision-based approaches to MAV stabilization, object tracking and autonomous landing. Finally, report experimental flight results for both the virtual testbed as well as for flight tests in the field, and discuss how algorithms developed in the virtual testbed were seamlessly transitioned to real flight testing.

We present a probabilistic framework--namely, multiscale generative models known as Dynamic Trees (DT)--for unsupervised image segmentation and subsequent matching of segmented regions in a given set of images. Beyond these novel applications of DTs, we propose important additions for this modeling paradigm. First, we introduce a novel DT architecture, where multilayered observable data are incorporated at all scales of the model. Second, we derive a novel probabilistic inference algorithm for DTs--Structured Variational Approximation (SVA)--which explicitly accounts for the statistical dependence of node positions and model structure in the approximate posterior distribution, thereby relaxing poorly justified independence assumptions in previous work. Finally, we propose a similarity measure for matching dynamic-tree models, representing segmented image regions, across images. Our results for several data sets show that DTs are capable of capturing important component-subcomponent relationships among objects and their parts, and that DTs perform well in segmenting images into plausible pixel clusters. We demonstrate the significantly improved properties of the SVA algorithm--both in terms of substantially faster convergence rates and larger approximate posteriors for the inferred models--when compared with competing inference algorithms. Furthermore, results on unsupervised object recognition demonstrate the viability of the proposed similarity measure for matching dynamic-structure statistical models.

In this paper, we develop a unified vision system for small-scale aircraft that not only addresses basic flight stability and control, but also enables more intelligent missions, such as ground object recognition and moving-object tracking. The proposed system defines a framework for real-time image feature extraction, horizon detection and sky/ground segmentation, and contextual ground object detection. Multiscale Linear Discriminant Analysis (MLDA) defines the first stage of the vision system, and generates a multiscale description of images, incorporating both color and texture through a dynamic representation of image details. This representation is ideally suited for horizon detection and sky/ground segmentation of images, which we accomplish through the probabilistic representation of tree-structured belief networks (TSBN). Specifically, we propose incomplete meta TSBNs (IMTSBN) to accommodate the properties of our MLDA representation and to enhance the descriptive component of these statistical models. In the last stage of the vision processing, we seamlessly extend this probabilistic framework to perform computationally efficient detection and recognition of objects in the segmented ground region, through the idea of visual contexts. By exploiting visual contexts, we can quickly focus on candidate regions where objects of interest may be found, and then perform additional analysis for those regions only. Throughout, our approach is heavily influenced by real-time constraints and robustness to transient video noise.

Currently popular feature extraction tools (e.g., Gabor, wavelet analysis) do not economically represent edges in images. As a step towards solving this problem, the wedgelet transform was recently proposed D.L. Donoho, [1999]; this transform provides nearly optimal representation of objects in the Horizon model, as measured by the minimax mean-squared error. However, there is no reason to assume that the components useful for representing pixel values must also be useful for discriminating between regions in an image. Thus, having the successful extraction of edges as our goal, we propose a novel image analysis method-namely, multiresolution linear discriminant analysis (MLDA). In MLDA, analogously to the wedgelet transform, we seek directions that are efficient for discrimination. The MLDA framework comprises the following components: the MLDA atom, dictionary, tree, graph, and MLDA-based algorithms. In this paper, we explain these components and demonstrate the powerful expressiveness of MLDA, which gives rise to fast geometrical-structure-analysis algorithms.

Recently, we have implemented a computer-vision based horizon-tracking algorithm for flight stability and autonomy in micro air vehicles (MAVs) [S. M. Ettinger et al., 2002]. Occasionally, this algorithm fails in scenarios where the underlying Gaussian assumption for the sky and ground appearances is not appropriate. Therefore, in this paper, we present a general statistical image modeling framework which we have use to build prior models of the sky and ground. Once trained, these models can be incorporated into our existing horizon-tracking algorithm. Since the appearances of the sky and ground vary enormously, no single feature is sufficient for accurate modeling: as such, we rely both on color and texture as critical features in our modeling framework. Specifically, we choose hue and intensity for our color representation, and the complex wavelet transform (CWT) for our texture representation. We then use hidden Markov tree (HMT) models, which are particularly well suited for the CWT's inherent tree structure, as our underlying statistical models over our feature space. With this approach, we have achieved reliable and robust image segmentation of flight images from on-board our MAVs as well as on more difficult-to-classify sky/ground images.

Sensor fusion and sensor integration is becoming an increasingly popular approach in dealing with complex sensor systems in autonomous mobile robots (AMR). However, the procedure for the sensor integration and sensor fusion is a non-trivial process. This paper presents a scenario based approach to sensor fusion based on the autonomous evolution of sensory and actuator driver layers through environmental constraints (AEDEC) [T.A Choi, 2002]. Using the scenario based approach, the programmer's work of creating a sensory driver will be eliminated by having the AMR learn the driver on its own. In the process of creating each scenario, sensor fusion is automatically implemented. If sensors change or even if the sensor configuration changes, the driver can be updated by having the AMR relearn the driver over again. Due to the tabular structure of the scenario based sensory drivers, malfunctioning sensors can not only be detected, but the driver can automatically adapt to the malfunctioning sensor in real-time. Furthermore, different AMR's trained using AEDEC architecture will have similar interpretations of its environment. This is guaranteed by having the AMR learn the driver in the same highly structured training environment. The behavioral coding is simplified by eliminating any reference to hardware dependent parameters. Finally, the level of abstraction and the consistency of the highly structured environment allows for coding portability.

We are working on mapping multi-channel neural spike data, recorded from multiple cortical areas of an owl monkey, to corresponding 3D monkey arm positions. In earlier work on this mapping task, we observed that continuous function approximators (such as artificial neural networks) have difficulty in jointly estimating 3D arm positions for two distinct cases-namely, when the monkey's arm is stationary and when it is moving. Therefore, we propose a multiple-model approach that first classifies neural spike data into two classes, corresponding to two states of the monkey's arm: (1) stationary and (2) moving. Then, the output of this classifier is used as a gating mechanism for subsequent continuous models, with one model per class. In this paper, we first motivate and discuss our approach. Next, we present encouraging results for the classifier stage, based on hidden Markov models (HMMs), and also for the entire bimodal mapping system. Finally, we conclude with a discussion of the results and suggest future avenues of research.

Substantial progress has been made recently towards designing, building and test-flying remotely piloted Micro Air Vehicles (MAVs) and small UAVs. We seek to complement this progress in overcoming the aerodynamic obstacles to flight at very small scales with a vision-guided flight stability and autonomy system, based on a robust horizon detection algorithm. In this paper, we first motivate the use of computer vision for MAV autonomy, arguing that given current sensor technology, vision may be the only practical approach to the problem. We then describe our statistical vision-based horizon detection algorithm, which has been demonstrated at 30 Hz with over 99.9% correct horizon identification. Next, we develop robust schemes for the detection of extreme MAV attitudes, where no horizon is visible, and for the detection of horizon estimation errors, due to external factors such as video transmission noise. Finally, we discuss our feedback controller for self-stabilized flight, and report results on vision-based autonomous flights of duration exceeding ten minutes.

Models of human control strategy (HCS), which accurately emulate dynamic human behavior, have far reaching potential in areas ranging from robotics to virtual reality to the intelligent vehicle highway project. A number of learning algorithms, including fuzzy logic, neural networks, and locally weighted regression exist for modeling continuous human control strategies. These algorithms, however, may not be well suited for modeling discontinuous human control strategies. Therefore, we propose a new stochastic, discontinuous modeling framework, for abstracting human control strategies, based on hidden Markov models (HMM). In this paper, we first describe the real-time driving simulator which we developed for investigating human control strategies. Next, we demonstrate the shortcomings of a typical continuous modeling approach in modeling discontinuous human control strategies. We then propose an HMM-based method for modeling discontinuous human control strategies. The proposed controller overcomes these shortcomings and demonstrates greater fidelity to the human training data. We conclude the paper with further comparisons between the two competing modeling approaches and we propose avenues for future research. © 2001 John Wiley & Sons, Inc.

We address the problem of transferring human control strategies (HCS) from an expert model to an apprentice model. The proposed algorithm allows us to develop useful apprentice models that incorporate some of the robust aspects of the expert HCS models. We first describe our experimental platform, a real-time graphic driving simulator, for collecting and modeling human control strategies. Then, we discuss an adaptive neural network learning architecture for abstracting HCS models. Next, we define a hidden Markov model (HMM) based similarity measure which allows us to compare different human control strategies. This similarity measure is combined subsequently with simultaneously perturbed stochastic approximation to develop our proposed transfer learning algorithm. In this algorithm, an expert HCS model influences both the structure and the parametric representation of the eventual apprentice HCS model. Finally, we describe some experimental results of the proposed algorithm.

Models of human control strategy (HCS), which accurately emulate dynamic human behavior, have far reaching potential in areas ranging from robotics to virtual reality to the intelligent vehicle highway project. A number of learning algorithms, including fuzzy logic, neural networks, and locally weighted regression exist for modeling continuous human control strategies. These algorithms, however, may not be well suited for modeling discontinuous human control strategies. Therefore, we propose a new stochastic discontinuous modeling framework, for abstracting human control strategies, based on hidden Markov models. In this paper, we first describe the real-time driving simulator which we have developed for investigating human control strategies. Next, we demonstrate the shortcomings of a typical continuous modeling approach in modeling a discontinuous human control strategy. We then propose an HMM-based method of modeling discontinuous human control strategies, and show that the proposed controller overcomes these shortcomings and demonstrates greater fidelity to the human training data. We conclude the paper with further comparisons between the two competing modeling approaches.

In the last few years, modeling dynamic human control strategy (HCS) is becoming an increasingly popular paradigm in a number of different research areas, such as the intelligent vehicle highway system, virtual reality and robotics. Usually, these models are derived empirically, rather than analytically, from real human input-output control data. As such, there is a great need to develop adequate performance criteria for these models, as few guarantees exist about their theoretical performance. It is our goal in this paper to develop several such criteria. In this paper, we first collect driving data from different individuals through a real-time graphic driving simulator. We then model each individual's control strategy through the flexible cascade neural network learning architecture. Next, we develop two performance measures for evaluating the resulting HCS models, one dealing with obstacle avoidance, the other with tight-turning behavior. Finally, we evaluate the relative skill of different HCS models through the proposed performance criteria.

Modeling the dynamic human control strategy (HCS) is becoming an increasingly popular paradigm in a number of different research areas, ranging from robotics to intelligent vehicle highway systems. Usually, HCS models are derived empirically, rather than analytically, from real human input-output data. While these empirical models offer an effective means of transferring intelligent behaviors from humans to robots and other machines, the models are not explicitly optimized with respect to potentially important performance criteria. We therefore propose an iterative algorithm for optimizing an initially stable HCS model with respect to an independent, user-specified performance criterion. We first collect driving data from different individuals through a real-time graphic driving simulator. Next, we describe how we model each individual's control strategy through flexible cascade neural networks. Once we have initially stable HCS models, we propose simultaneously perturbed stochastic approximation (SPSA) to optimize these models with respect to a chosen performance criterion. Finally, we describe and discuss some experimental results with the proposed algorithm.

Modeling dynamic human control strategy (HCS), or human skill in response to real-time sensing is becoming an increasingly popular paradigm in many different research areas. We propose a stochastic similarity measure, based on hidden Markov model analysis, capable of comparing and contrasting stochastic, dynamic, multidimensional trajectories. We first derive and demonstrate properties of the similarity measure for stochastic systems. We then apply the similarity measure to real-time human driving data by comparing different control strategies among different individuals. We show that the proposed similarity measure out performs the more traditional Bayes classifier in correctly grouping driving data from the same individual. Finally, we illustrate how the similarity measure can be used in the validation of models which are learned from experimental data, and how we can connect model validation and model learning to iteratively improve our models of HCS.

Modeling dynamic human control strategy (HCS), or human skill through learning is becoming an increasingly popular paradigm in many different research areas, such as intelligent vehicle systems, virtual reality, and space robotics. Validating the fidelity of such models requires that we compare the dynamic trajectories generated by the HCS model in the control feedback loop to the original human control data. To this end we have developed a stochastic similarity measure-based on hidden Markov model (HMM) analysis-capable of comparing dynamic, multi-dimensional trajectories. In this paper, we first derive and demonstrate properties of the proposed similarity measure for stochastic systems. We then apply the similarity measure to real-time human driving data by comparing different control strategies for different individuals. Finally, we show that the similarity measure outperforms the more traditional Bayes classifier in correctly grouping driving data from the same individual.

Modeling dynamic human control strategy, or human skill, in response to real-time sensing is becoming an increasingly popular paradigm in many research areas. These models are learned from experimental data, and as such can be characterized despite the lack of a good physical model. Unfortunately, learned models presently offer few, if any, guarantees in terms of model fidelity to the source data. As such, we propose an independent, post-training model validation procedure based on hidden Markov models (HMMs). The proposed method generates a stochastic similarity measure comparing system trajectories for the source process and the learned models. Using this method, we are able to verify model fidelity. We demonstrate the proposed method in the validation of neural-network models for real-time human driving skill.

Much work in recent years has focused on transferring human skill to robots by abstracting that skill into a machine-understandable, computational model. Such skill models, however, can be used not only for transferring human control strategy to robots, but also for helping less-skilled human operators improve their performance. The authors propose a two-step approach for transferring skill from human expert to human apprentice. An expert's relevant control strategies or skills are first abstracted into a sensory-based computational model. Afterwards, this trained computational model is used to generate on-line advice for less-skilled operators who need to improve their skill. This advice can take advantage of many different sensor modalities, thereby potentially improving both the quality and speed of learning for the apprentice. Furthermore, this approach allows for the efficient transfer of skill from a single expert to many apprentices, as well as from many experts to a single apprentice. In this paper, the authors first describe a flexible neural-network-based method for modeling human control strategy and provide motivation for its use. The authors then present a case study for teaching control strategy from one person to another in this two-step approach of transferring skill.

The self-mobile space manipulator (SM/sup 2/) has evolved to adapt to the new pre-integrated I-beam structure of the Space Station Freedom (SSF). In this paper, we first briefly overview the update of the robot configuration and testbed. The new robot is capable of projecting cameras anywhere interior or exterior of SSF, and will be an ideal tool for inspecting connectors, structures, and other facilities on SSF. Experiments have been performed under two gravity compensation systems and a full-scale model of a segment of the SSF. This paper then presents a real-time shared control architecture that enables the robot to coordinate autonomous locomotion and teleoperation input for reliable walking on SSF. Autonomous locomotion can be executed based on a CAD model and off-line trajectory planning, or can be guided by a vision system with neural network identification. Teleoperation control can be specified by a real-time graphical interface and a free-flying hand controller. SM/sup 2/ will be a valuable assistant for astronauts in inspection and other EVA missions.>

Inverse kinematics is computationally expensive and can result in significant control delays in real time. For a redundant robot, additional computations are required for the inverse kinematic solution through optimization schemes. Based on the fact that humans do not compute exact inverse kinematics, but can do precise positioning for heuristics, an inverse kinematic mapping using fuzzy logic is developed. The implementation of the scheme has demonstrated that it is feasible for both redundant and nonredundant cases, and that it is very computationally efficient. The result provides sufficient precision, and transient tracking error can be controlled based on a fuzzy adaptive scheme proposed in the paper.