If a wheelchair for disabled is used for semi-autonomous navigation indoors it must be able to navigate through doors. A door and doorway can be parameterized with five parameters. A divide and conquer implementation of the Hough transform is used to segment outlines from range scans. The client software remote controls the wheelchair from the MATLAB environment. The software consists of several Java threads that run concurrently. Sensor data are polled by threads and put into databases to reduce the network lag. The databases are used by a controller and a Kalman filter. Since most of the implementation is coded in Java it is possible to run it as a stand alone program on a computer that has Java installed. From 10 runs the trajectory offset was calculated to 0.9 cm with a standard deviation of 1.4 cm. The standard deviation of the heading was 2.2 degrees. This performance is essentially independent of the initial starting pose.

This paper presents an integrated, low-level approach to removing sensor noise, cross talk, spurious specular reflections, and solving the association problem in a light stripe scanner. Most single-camera scanners rely on the laser brightness exceeding that of the entire image. Our system uses two cameras to measure the stripe and combines the knowledge of the light plane orientation to produce useful validation properties. The key result is the development of a condition relating image plane measurements and camera intrinsic parameters, which allows validation/association to be performed independently of 3D reconstruction. The same equations are used to improve ranging accuracy compared to single-camera systems. We also show how the system may be self-calibrated using measurements of an arbitrary nonplanar target. As validation allows the operation in ambient light, the registered colour and range are captured in the same sensor. An experimental scanner demonstrates the effectiveness of the proposed techniques.

The problem studied is feedback for docking a mobile robot or AGV, to a pallet. The pallet is of known size but with an essentially unknown load. The pallet has an initial uncertainty in pose (position and orientation) of the order /spl plusmn/15 cm and /spl plusmn/20 degrees. The docking error is required to be within /spl plusmn/1 cm and /spl plusmn/1 degree with "very low" failure rate. For the docking a combination of a range camera and a video camera is used. In the paper the range camera is emphasized. Experimental results from this work in progress are presented. Successful docking has been made with typical /spl plusmn/5 mm as errors. Currently one weak part is the integration with the control system on board the robot. Our persistent experience from this and earlier tests is that the weak part when using non-contact sensing for feedback in robots is the association problem. It should be mentioned that the resolution of a range camera is strongly distance dependent. One finding in the paper is that this type of docking is feasible and can be made self-monitoring.

Consider a robot to measure or operate on man made objects randomly located in the workspace. The optronic sensing onboard the robot are a scanning range measuring time-of-flight laser and a CCD camera. The goal of the paper is to give explicit covariance matrices for the extracted geometric primitives in the surrounding workspace. Emphasis is on correlation properties of the stochastic error models during motion. Topics studied include: (i) covariance of Radon/Hough peaks for plane surfaces; (ii) covariances for the intersection of two planes; (iii) equations for combining vision features, plane surfaces and range discontinuities; and (iv) explicit equations of how the covariance matrices are transformed during the robot motion. Typical applications are; models for verification and updating of CAD-models when navigating inside buildings and industrial plants, and accumulating sensor readings for a telecommanded robot.

This paper is on sensor based control for guiding a robot to correct gripping of objects having a large position uncertainty. An eye-in-hand mounted range camera is considered. A probabilistic problem formulation based on the requested posture at gripping and corresponding tolerances is presented. The problem is solved approximately using dynamic programming for a 1-degree-of-freedom manipulator. A five-step dual control law is studied in more detail. A typical case is that in the first part of the control sequence the robot steers towards the optimal sensing position and in the last part the error with respect to the gripping posture is minimized. Since range camera sensing introduces both range dependent noise and occlusion there is a need for 'exploratory moves'. This behavior is formalized and includes 'dual control'.

This paper describes two related tests; to navigate an articulated lawn mower and to reverse a mobile robot with a trailer. In both cases the vehicles are to follow a prespecified trajectory. The navigation principal is based on measured directions to several identical beacons, consisting of strips of reflective tapes. The angular sensor is a rotating laser for the illumination of the beacons and a highly sensitive electro-optical receiver for detecting the directions to the beacons. A Kalman filter is used to combine the measurements from the odometers with the detected angles to the known position of the beacons. To measure the angle between the robot and the trailer the same laser was used. This was done by placing two reflective beacons on the trailer. The repeatability was within 2 centimetre at low speed. The navigation of these two different types of vehicles turns out to be, essentially, the same problem. The sensitivities are different. Emphasis is on robust state estimation.>

This paper considers detection of errors during assembly. The sensors used are a combination of accelerometers, a force/torque (F/T) sensor; and an electrical contact indicator. Preliminary experiments show that by combining the three types of sensing, both normal contact situations and defects on the parts can be identified. For example, ruts were not detected by the F/T sensor but by the accelerometer. On the other hand the acceleration impulse from a peg with a burr was weak while the F/T sensor combined with the electrical contact indication gave distinct output on the matter. The contact indicating current was fluctuating just after an impact, which shows how the peg bounces on the surface of the chamfer or hole. Using the full range of sensors in assembly of metal workpieces the nature of the error can be determined.>

Consider a wheeled robot/vehicle equipped with odometers, rate gyros and accelerometers. The problem studied is to extract geometric properties of the surface using these sensors. Same basic results are: conditions are given on the curvature for the point of contact to move continuously along the supporting surface. An robust approximate algorithm for reconstructing a piecewise continuous second order surface is given, where only a rate gyro is used. A 7 meter long ramp, consisting of segments of plane surfaces, was reconstructed within +/-2 cm using a rate gyro with 5 degrees//spl radic//sub (hr/) in drift. Experimental results, when the robot traverses surfaces with different degrees of discontinuity, are presented. The test surfaces included steps and ramps with different heights and inclinations.>

The long term goal is to find generic methods for navigating a robot relative to objects of arbitrary shape. Both range measuring lasers and range cameras are considered. The authors study the distance transform on measurements for estimating the motion of a robot relative to an object of arbitrary shape. Emphasis is on the error mechanism when tested on cluttered outdoor and indoor scenes. Also the laser measurements are modelled for different types of object/surface properties. Potential applications are outdoor vehicles as well as industrial robots and AGVs.>

A method to determine gripper parameters in situ is given, and first results from model tests in plane motion are presented. Robotic assembly is associated with impacts on the gripper. These may be modeled to find contact force position and line of action in the gripper, relative orientation between gripper and contacted surface and conditions at the point of insertion. During impact the acceleration pole maps contact point kinematics which may be related to the conventional impact diagram. Kinematic modeling, dynamic and geometric modeling, and the impact diagram are discussed. The analysis shows that the point of application of the impacting force and its line of action may be found from measured accelerations.>

The problem addressed is sensor feedback for guiding a robot to the correct gripping point and to arrive within the prescribed tolerances. The solution is divided into three parts, the extraction of posture parameters from a range camera dynamic filtering to reduce spurious responses and a linear quadratic Gaussian designed feedback law to guide the robot to the correct gripping point. The prescribed tolerances were reached within a reasonable time. For transparent results a mobile was considered with three state variables. This is a first simple case toward a unified theory of tolerance controlled gripping.>

The goal behind this paper is to find a generic method for controlling the motion of a robot relative to an object of an arbitrary shape. In this paper we study; - modelling laser range measurements for different type of objectslsurface properties. Outdoor scenes are emphasized. - testing the distance transform on measurements for estimating the motion of a robot relative to an object of arbitmy shape. Emphasis in the present study is to get experience of the error mechanism in the distance transform when tested on cluttered laser measurements. Both natural and man-made objects are used and compared in the tests.