Image based visual servoing (IBVS) has become increasingly common in automated systems. IBVS control systems are dependent on the accuracy of an image Jacobian, which requires the knowledge of the internal and external parameters of the system. In most systems, some or all of these parameters are unknown. A centralized motion algorithm (CMA) is proposed to compute the image Jacobian with fast convergence and low iterations. Centralized motion is defined as a motion where a single feature point does not change its pixel coordinate value through the camera's motion. Therefore, the CMA is an algorithm that exploits the fact that the chosen feature point will have little to no changes in pixel coordinates. As a result, the CMA requires only a single stationary feature point, along with a minimum of two more feature points that will be dynamic, in the image to compute the image Jacobian. Using a Whole-Arm Manipulation (WAM) IBVS system, it was shown that an image Jacobian with an average condition number of 45, converges in an average of 4 iterations from 6 different starting positions. Experiments show that the CMA is a fast and effective online calibration algorithm.

BECCA (a Brain-Emulating Cognition and Control Architecture software package) was developed in order to perform general reinforcement learning, that is, to enable unmodeled embodied systems operating in unstructured environments to perform unfamiliar tasks. It accomplishes this through automatic paired feature creation and reinforcement learning algorithms. This paper describes an implementation of BECCA on a seven Degree of Freedom (DoF) Barrett Whole Arm Manipulator (WAM) undergoing a series of experiments designed to test the reinforcement learner's ability to adapt to the WAM hardware. In the experiments, the following is demonstrated, 1) learning to transition the WAM between states, 2) learning to perform at near optimal levels on one, two and three dimensional navigation tasks, 3) applying learning in simulation to hardware performance, 4) learning under inconsistent, human-generated reward, and 5) combining the reinforcement learner with Probabilistic Roadmap Methods (PRM) to improve scalability. The goal of the paper is to demonstrate both the scalability of the BECCA reinforcement learning approach using different formulations of the state space and to show the approach in this paper operating on complex physical hardware.

In this paper we relax the assumption of network connectivity within the sensor network and introduce mobile communication relays to the network. This addition converts the homogeneous sensor network to a heterogeneous one. Based on the communication geometry of both sensing and communication relay agents we derive communication constraints within the network that guarantee network connectivity. We then define a heterogeneous proximity graph that encodes the communication links that exist within the heterogeneous network. By specifying particular edge weights in the proximity graph, we provide a technique for biasing particular connections within the heterogeneous sensor network. Through a minimal spanning tree approach, we show how to minimize communication links within the network which allows for larger feasible motion sets of the sensing agents that guarantee the network remains connected. We also provide an algorithm that allows for adding communication links to the minimal spanning tree of the heterogeneous proximity graph to create a biconnected graph that is robust to a single node failure. We then combine a prioritized search algorithm and the communication constraints to provide a decentralized prioritized sensing control algorithm for a heterogeneous sensor network that maintains network connectivity.

The twin-pulse motion profile is still perhaps the simplest profile in existence for reducing residual vibration. Its primary drawback is that of narrow bandwidth, i.e., it is effective for attenuating only a very narrow range of frequencies about its design frequency. This has led to the inclusion of more pulses to improve robustness to frequency variation. However, one cost of adding pulses is extending the duration of the resulting motion. This paper presents several techniques whereby the basic twin-pulse motion profile can be expanded to reduce its frequency sensitivity while still preserving its advantages. These techniques include intentionally mis-aligning the null points of ideal twin pulses and a convolved boxcar function, then adding a trapezoidal profile component, finally culminating in what we term a misaligned boxcar/trapezoidal/twin pulse (M- BC:TZ:TP) motion profile. This expanded profile and a basic twin-pulse profile were both evaluated on a ten-pendula system with natural frequencies varying over a range of nearly 10%. Simulation results showed an overall reduction of 97% with the expanded twin-pulse profile, and notably all ten frequencies were attenuated; only one frequency was attenuated with the standard twin-pulse profile. Experiments using a linear rail transport system and the ten pendula device validated the simulation. These expanded profiles retain the simplicity and short duration of the twin-pulse with increased robustness to frequency variations.

Motion profiles such as swing free and input shaping are intended to eliminate residual vibrations. This paper demonstrates a simple generalized method of creating and understanding pulse-based profiles, and presents an approach to suppress a wider band of induced vibration than previously attainable. The current approach of input shaping ignores the potential contributions the base profile and pulse shape can have on the profile. When examined from a signal processing point of view these pulse-based profiles can be thought of as sampled low pass filters. The sampling of the filter kernel defines the magnitude of the pulses, but as in any digital system it also produces aliasing. An interpolating filter can then be used to both remove the aliasing and to smooth and shape the base profile. This dual purpose of the interpolating filter allows the use of the swing free step input in conjunction with a rate-limit as a base profile. A generalized methodology is proposed which can create smooth motion profiles with continuous derivatives, while keeping the computational efficient of traditional input shaping. Experimental validation of two multi-DOF systems is shown.

This paper describes a method that requires only a single projected point to calibrate a pair of cameras attached to pan-tilt units (PTUs) in a hand-eye robot configuration. Most existing calibration methods require either a set of known calibration locations or a large number of corresponding image features. These requirements usually imply human intervention, which complicates the calibration procedure. Rather than using multiple points in the workspace, the proposed algorithm uses a single stationary point generated by aiming a laser at a planar surface for the entire calibration. The single point provides a high intensity feature that is easy to extract from the image. To exercise all of the PTU and camera degrees of freedom, a large number of random robot/PTU configurations are computed automatically and 60 configurations, where both cameras can extract the single calibration point, are retained. As a result, this approach minimizes image processing requirements and virtually eliminates the correspondence problem. Using the Levenberg-Marquardt non-linear minimization algorithm, the Denavit-Hartenberg (D-H) parameters for the pan-tilt axes along with the internal and external parameters of the cameras are computed. This method requires no human intervention, employs simple image processing to detect the high intensity laser point, and produces a triangulation error of roughly 3mm at a range of 1m.

Automated transport of suspended objects is a subject of importance in many manufacturing, construction, and military applications. Suppression of the natural oscillation of payloads after a transport motion has been extensively studied, but generalized planar motion has yet to be examined. Obstacles in a crane or robot workspace may necessitate transport using a sequence of many linear segments or by a sequence of fewer curvilinear ones. The use of curvilinear motions in such cases may have the following advantages: (1) less error buildup in optimization due to the use of fewer segments, and (2) faster transport. We investigate parametrically-defined polynomial spatial paths optimized using dynamic programming. We present simulations and experimental evaluations of these optimizations.

This paper describes a robot calibration approach called the virtual closed kinematic chain (ViCKi) method. Traditionally, calibration requires the measurement of the position and orientation of the end effector, and measurement resolution limits the accuracy of the robot model. In ViCKi, we attach a laser to the end effector to create a virtual 7th link. The laser spot produced on a distant plane, the end of this virtual link, magnifies small changes at the end effector, resulting in a high resolution error measurement of the end effector. The accuracy of the robot after using the proposed calibration procedure is measured by aiming at an arbitrary fixed point and measuring the mean and standard deviation of the radius of spread of the projected points. The mean and standard deviation of the radius of spread were improved from 5.64mm and 1.89mm to 1.05mm and 0.587mm respectively. It is also shown in simulation that the method can be automated by a feedback system that can be implemented in real-time.

This paper describes an industrial robot calibration algorithm called the virtual closed kinematic chain method. Current robot kinematic calibration methods use measurements of position and orientation of the end effector. The accuracy of these measurements is limited by the resolution of the measuring equipment. In the proposed method, a laser pointer tool, attached to the robot's end effector, aims at a constant but unknown location on a fixed object, effectively creating a virtual 7 DOFs closed kinematic chain. As a result, small variations in position and orientation of the end effector are magnified on the distant object. Hence, the resolution of observations is improved, increasing the accuracy of joint angle measurements that are required to calibrate the robot. The method is verified using both simulation and real experiments. It is also shown in simulation that the method can be automated by a feedback system that can be implemented in real time. The accuracy of the robot after using the proposed calibration procedure is measured by aiming at an arbitrary fixed point and measuring the mean and standard deviation of the radius of spread of the projected points. The mean and standard deviation of the radius of spread were improved from 5.64 and 1.89 mm to 1.05 and 0.587 mm, respectively.

In the high speed transport of freely suspended objects, traditional methods of suppressing residual oscillations may not produce satisfactory results. Nonlinearities may exist to the point that the linearizing assumptions required by these methods may no longer be valid. The optimization algorithm of dynamic programming (DP) is advantageous for such intrinsically nonlinear applications. This paper will outline the development of the DP algorithm for discrete time systems. It also describe the application of the algorithm to a rapid maneuver of a long slender payload suspended by a cable at each end. The cables are each held by one of two independent robot manipulators. The DP algorithm will be used to generate optimal swing-free motion trajectories for the robots. Simulation and experimental results show that residual oscillations are virtually nonexistent

The topic of composing inputs for systems with mechanical flexibility has received attention for many years. Two popular methods are (1) shaping a nominal trajectory by convolution with an impulse sequence which itself cancels the flexibility, and (2) trajectory parameter optimization using structured basis functions, e.g. bang-coast-bang. However, each of these methods deteriorates in a different way when the motion is rapid compared to the natural period of the flexibility. Motions synthesized using dynamic programming offer clear advantages in these situations: (1) the resulting motions are smoother than those using impulse-convolution, and (2) dynamic programming is deterministic and does not rely on a convex objective function like gradient-based parameter optimization. Examples of the shortcomings of impulse-convolution and parameter-optimization will be shown; then we present both simulation and experimental evaluation of rapid transport of a suspended object using a robot manipulator and a minimum-energy trajectory obtained by dynamic programming.

With the rapid development of personal computers (PCs) in the past 15 years, a trend has emerged in which conventional industrial microcontrollers are being replaced by PCs because they provide more computation power while at the same time support a user-friendly interface for better interoperability. Although there are many research efforts in this direction, a few key issues concerning the determinism of the PC hardware and firmware were overlooked that have critical impact on the hard real-time performance of such systems. Key issues include: the nondeterministic occurrence of system management interrupts (SMI), the nondeterministic I/O latency caused by PCI bus arbitration and hard disk buffering. These issues along with suggested solutions will be addressed in this paper. Due to the diversity of firmware (BIOS) and hardware design from different PC manufacturers, it is impossible to make a comprehensive list of rules to safeguard a deterministic controller PC configuration. The goal of this paper is therefore to present a general design guideline that can be used to analyze and configure a specific system with the help of necessary motherboard/chipset/BIOS manuals.

Several methods are available for generating input trajectories that eliminate residual vibrations for a variety of systems. These methods have been successfully applied to an assortment of both linear and non-linear situations. While multiple input systems have been addressed, the issue of developing residual vibration-suppressed input trajectories for cooperative manipulators has not been satisfactorily covered in the literature. This research focuses specifically on the problem of multiple cooperative manipulators working in conjunction to move a common payload. A method is determined by which multiple modes of residual vibration are suppressed, and an input trajectory is computed for dual manipulators which reduces residual vibrations by at least seventy-five percent.

In this paper we propose a cooperative motion scheme that is able to control the sharing of gravitational and inertia load (forces required for load acceleration/deceleration) among multiple position-controlled robots. We adopt a clamped-free configuration, where a leader robot grasps the load directly and a follower robot grasps the load through a spring mounted on its end effector. While the leader travels the designated trajectory, a delta trajectory is added to the follower that supplies its share of inertia force by stretching/compressing the spring. The trajectory of the follower is pre-shaped to cancel out vibrations, therefore reducing the impact force on the robot mechanism. The spring also provides compliance in non-force-controlled directions to reduce internal forces. We analyze the influence of temporal error on both load sharing and internal force and propose a synchronization method to reduce this error to less than 1 ms. Every aspect of our scheme can be implemented with commercial robots without any modification of controller parameters or knowledge of the manipulator dynamics. Being able to execute our cooperative motion at a speed of 0.5 m per sec, it is much faster than impedance/hybrid control based approaches, suggesting the ability to satisfy the cycle time requirement for industrial applications such as material handling.

This paper presents an efficient method to grasp an unknown planar object (disk-like object) using a three-fingered Barrett band. First, active contour models are extracted from the images taken by Biclops, a dual camera directed vision system. The centroid of each active contour model is computed, employed in triangulation, and used as an estimate of a 3-D centroid point of the object. To compute precisely where the gripper will go and how to configure the gripper to grasp the object, the active contour models are also used in a polar coordinate frame to extract concave and convex sections of the model. The proposed algorithm attempts to place the thumb in concave sections of the object, with the remaining two fingers place such that a grasp quality metric is maximized. As a consequence, the geometric properties of the object itself can improve the quality of grasp. In this work, the gripper contact is modeled as a surface rather than a point. This makes the interaction between the object and the gripper more realistic. Only a small set of inflection points are chosen as potential grasping locations. Thus the complexity of the proposed algorithm is linear as opposed to cubic if all the points on the object boundary are chosen as potential grasping locations.

The use of two or more robots jointly manipulating a common object, termed cooperative control, has many advantages in industrial applications. Much of the previous work on cooperative control, however, has emphasized capabilities, such as joint torque control, not generally present on industrial robots. This paper applies the concept of cooperative control to industrial robots by taking advantage of those capabilities that do exist on industrial robots. Specifically, a design approach that makes use of tool-based coordinate systems, trajectory generation, real-time modification, and distributed control of multiple robots is presented. These concepts are implemented on a system of two industrial robots using an adaptive fuzzy controller, and the results are discussed.

A new strategy to select observer (camera) viewpoints for global 3D reconstruction of unknown objects is presented. The method has four steps: local surface feature extraction, shape classification, viewpoint selection and global reconstruction. An active vision system (Biclops) with two cameras aimed by independent pan/tilt axes, extracts 2D and 3D surface features from the scene. These local features are assembled into simple geometric primitives. The primitives are then classified into shapes, which are used to hypothesize the global shape of the object. The next viewpoint is chosen to verify the hypothesized shape. If the hypothesis is verified, some information about global reconstruction of a model can be stored. If not, the data leading up to this viewpoint is re-examined to create a more consistent hypothesis for the object shape. The paper has two main contributions. First, the next viewpoint algorithm uses only the local geometric features of an object. Second, the visibility constraint is not used in the function to compute next viewpoint. Instead it is solved prior to viewpoint selection using volume intersection of prismatic cones generated from camera coordinate center and image plane. The proposed algorithm is demonstrated experimentally by reconstructing the model for simple 3D objects using a two-camera stereo vision system mounted on a 6-DOF manipulator in an uncontrolled (noisy) environment.

This paper addresses the long-standing problem that visual feedback delay can drive a high bandwidth visual servoing system unstable. While some architectures become unstable with increasing image processing delay, we propose a novel visual servo implementation whose stability is independent of delay time. Performance decreases as image processing delay increases, but the system always remains stable. We demonstrate our new approach experimentally on a camera-pointing device. For those visual servo implementations that will become unstable with increased time delay, we propose an analytical method to find the upper limit of acceptable delay, using equations derived from the Nyquist stability criterion. The approach described in this paper is not only a useful tool for design and implementation of stable visual servoing systems, but also has general application to other dynamic systems with delay elements.

We investigate "intelligent" grasping schemes using a fuzzy logic rule base expert system. We use a vision system, robot arm and mechanical hand to locate and manipulate unmodeled, randomly placed objects of various sizes and shapes. In the pregrasp stage, we use vision data to provide a nonlinear mapping from object characteristics to hand configuration. In the postgrasp stage, we use hand data to ascertain the security of the grasp. Computational geometry is used to gauge the quality of the grasp and to quantify and validate the choice of hand configurations generated by the fuzzy logic expert system. The system is implemented within a low-cost virtual collaborative environment.

The Department of Energy (DOE) in the USA, in its effort to clean up and shut down many of its radioactive facilities, would like to use robots to enhance efficiency and limit worker exposure to hazardous substances. Unlike industry, where the work environment is designed to facilitate robots, the DOE applications often have unknown, unstructured work areas. The paper describes an approach to use a trinocular image processing system to build a real time, simple representation of the objects in the workspace. This representation is needed for safe navigation as well as for manipulation of the objects in the workspace.

This paper presents the design, implementation, and use of a cycle time reduction system (CTRS) for remote cleanup and remediation of hazardous waste sites. The goal is to reduce the amount of time between a site-defined need to a site-delivered implementation of the robotic hardware that solves the problem. The CTRS uses simulation, followed by direct conversion to actual hardware, to meet the requirements of the application, and populates the solution with models of robots, grippers, and other components. The user can easily change robots, grippers, vision systems, control algorithms, sensors, or any other component in the current solution to attempt to find a more effective solution. The CTRS simulates the overall performance of the system. The interfaces between simulated components are the same as the physical interfaces between real components, thereby performing a simulated system integration of both the mechanical and data linkages. In this way, the CTRS will enable expedient transformation of the simulation to actual hardware operation. Once the solution is selected, the simulation is used to train an operator, while the procurement and integration of the real hardware components takes place. Upon integrating the hardware, the system is delivered to the site with a trained operator ready to operate the system.

The problems associated with the development of robots for space applications are reviewed. Within the Space Station project context, robots will operate in a relatively unstructured environment and must have the ability to deal with unexpected events. For this to be possible, a great deal of experimentation is required to determine the best algorithms for task decomposition, world modeling, and sensory processing to be able to predict their utility for particular tasks. An architecture which supports the evolutionary development of the robot is presented. The architecture forms the basis for a testbed system which allows researchers to develop, test, and evaluate different hardware and software approaches. The same architecture is also used in the telerobot, so that the transition from the teleoperated mode to more autonomous modes of robot operation is one of gradual rather than abrupt change. This architecture, the NASA/NBS standard reference model for telerobot control systems, has been adopted by NASA for the Flight Telerobotic Servicer, a two-armed robot intended to help build and maintain the Space Station.>

A description is given of a low-cost laser ranger suitable for robot applications. The region to be observed is illuminated by a sinusoidal intensity distribution generated by an optical projection system with a laser source. The deformation of the sine-wave pattern is detected by a phase-sensing method widely used in optical interferometry, and can be easily related to the shape and range information. Over a range determined by the depth of field of the camera at a given aperture and focal setting, high-resolution measurements can be made. Experimental results show that at a camera distance of about 1 m, the resolution and repeatability in locating a flat 25-mm/sup 2/ surface are of the order of 0.2 mm.>

A hidden feature elimination algorithm based on ray tracing through an octree is presented. The algorithm has been used in a model-based vision system for a robot which services numerically controlled machine tools in an automated factory. It determines which features of an object are visible from a given camera location. This prediction is used by the model-based vision system to keep the 3-D world model as close as possible to reality. Finally the effect of the resolution limitations of the octree on the algorithm is discussed.

The Sensory-Interactive Robotics Group at the National Bureau of Standards is producing PIPE, a pipelined image-processing engine, for research in low-level machine vision. PIPE processes sequences of images at field rates through a series of point and neighborhood operations. It is divided into a variable number of identical stages, each of which performs an independent set of operations on the image data stored in the stage. A stage control unit determines the sequence of operations performed within a stage on each image. This sequence is easily modified by a host computer during the inter-field interval when all of the stage control units can be totally reconfigured. Images flow through PIPE in several ways. In addition to the (standard pipeline) "forward" pathway, where an output image is sent to the next stage, an output image can also be sent to the same stage via a "recursive" pathway and to the previous stage via a "retrograde" pathway. As a result, PIPE can support relaxation operations, temporal neighborhood operations, and other local operations. Several processing modes are available in PIPE in addition to the usual "SIMD" mode of pipelined processors. In an "MIMD" mode, one of several operations is performed on a region of interest which can be defined by the host device or by previous image operations. PIPE also supports variable resolution pyramids where an image is compressed or expanded as it passes between stages.