This paper describes the functionality of a local planner module for autonomous mobile robots. As an intermediate layer of a hierarchical planning system for autonomous vehicles, this module translates goals provided by a map-based planner into general vehicle action modes called activities. Activities are composed of pre-defined sets of reflexive behaviors which yield known performance characteristics under certain specific environmental conditions. Activity selection is performed on the basis of perception-supplied descriptions of the local environment, descriptions of expected terrain characteristics from a map-based planner, local map information accumulated by the local planner itself, and knowledge stored with each activity describing how that activity may achieve various goals and handle various failures. Local planner reasoning occurs within a monitoring process associated with the selected activity. Selected activities communicate with the local planner by posting messages onto a blackboard. These messages convey information to the local planner such as the distance to the nearest obstacle in a certain direction and activity completion status. As the vehicle moves, the local planner builds a map that records the path traversed, as well as major features and landmarks encountered. This information is used when the vehicle needs to backtrack. The local planner is implemented in Lisp, and has been demonstrated in a simulated environment.

This paper presents a report on the WPI autonomous mobile robot (WAMR). This robot is currently under development by the Intelligent Machines Project at WPI. Its purpose is to serve as a testbed for real-time artificial intelligence. WAMR is expected to find its way from one place in a building to another, avoiding people and obstacles enroute. It is given no a priori knowledge of the building, but must learn about its environment by goal-directed exploration. Design concepts and descriptions of the major items completed thus far are presented. WAMR is a self-contained, wheeled robot that uses evidence based techniques to reason about actions. The robot builds and continually updates a world model of its environment. This is done using a combination of ultrasonic and visual data. This world model is interpreted and movement plans are generated by a planner utilizing uses real-time incremental evidence techniques. These movement plans are then carried out by a hierarchical evidence-based adaptive controller. Two interesting features of the robot are the line imaging ultrasonic sensor and the video subsystem. The former uses frequency variation to form a line image of obstacles between one and twenty feet in front of the robot. The latter attempts to mimic the human eye using neural network pattern recognition techniques. Several items have been completed thus far. The paper describes some of these, including the multiprocessor navigator and non-skid motion control system, the ultrasonic line imager, the concepts of the vision system, and the computer hardware and software environment.

The autonomous land vehicle (ALV) has successfully navigated simple non-branching roads. To traverse more complicated road-ways, the ALV must be able to detect road intersections. We present two methods of road intersection detection. The first recognizes an image road intersection by matching image road boundary data with a predefined intersection model. The predefined model is the intersection configuration derived from an a priori map database. Road boundaries are extracted from the image and processed to produce a set of linear segments. A heuristic method produces a set of combinations of these segments that correspond to segments in the model. The best match is determined by a minimum goodness-of-fit measure. The method was tried on a Y-intersection. The ALV uses appropriate models derived from a map database. A second method, rather than recognizing the image road intersection, finds corridors of free space that the ALV can navigate through. A road height profile obtained from a range scanner is used to find a free space region. The height profile, representing height and normal road surface angle, is broken into solid horizontal planar segments. In a segmented image, distances to the road boundary in each of five principal directions (east, northeast, north, northwest, and west) are found for each road point. These image distances are converted to distances between the road point and the world plane boundary in each direction using the road height profile. The smallest directional distance approximates the world road boundary distance. We present results on synthetic and real road intersection images.

An intelligent mobile robot must be able to accept a mission statement and constraints, plan its actions, execute its plans, perceive and adapt to its environment, and report its successes and failures. In this paper we describe a modular system architecture for such a complex mobile robot system. On-board versus off-board processing is a key system-level issue. We have selected off-board processing because the anticipated computer quantity, size, power requirement, and lack of robustness made on-board processing impractical if not impossible. Our system includes a transportable command center and a computer-controllable M113 armored personnel carrier, our mobile robot. The command center contains communication and computer hardware necessary for receiving and processing robot motion and sensor information, and for generating and transmitting mobility and sensor commands in real time to the robot. All control and status data transmission, between the robot and the command center, is accomplished through microwave links using a wide band, auto-tracking antenna. Under development since 1982, this system has demonstrated the capability of mission and route planning with execution at 8 km/hr, obstacle detection and avoidance at 15 km/hr, autonomous road following at 24 km/hr, and a remotely managed route reconnaissance mission at vehicle speeds of up to 40 km/hr.

Range imagery from a laser range imager developed at ERIM to support the DARPA ALV program provides a robust source of information to direct autonomous vehicle movement. Forward looking range imagery is processed into plan view map information of the surrounding environment for guidance and identifying obstacles. Intensity data from the sensor also provides a correlated object reflectivity image. The image processing required to transform the forward looking range data into top down format, and to subsequently find the vehicle path is completed at a nominal two second frame rate. If and when image processing speed is comparable to sensor rate (two frames per second), this would support a vehicle speed of approximately 40 miles per hour. This processing rate is possible using the current generation of Cytocomputer ® technology and can be further improved by migrating more of the image processing algorithm from a host VAX to the dedicated Cytocomputer processing system.

This paper describes range image based obstacle avoidance methods developed to run on the Warp Machine, a new linear systolic array processor capable of delivering up to 100 million floating point operations per second. The Warp processing described here has been tested on the DARPA Autonomous Land Vehicle (ALV) to detect and locate obstacles present in an outdoor road environment at speeds up to 10 km/hour. The overall perception system consists of two subsystems; a color image processing subsystem used to find the road area and the Warp based range image processing subsystem. We briefly describe the architecture of the Warp Machine and methods of exploiting the parallelism of the Warp system. Next we describe the details of the obstacle avoidance perception algorithm and its mapping onto the Warp Machine. Results of applying the system in an experimental test run are presented.

Following a brief review of multi-sensor image processing techniques for obstacle detection, we consider a new method to employ range data to extract object regions of interest from an outdoor natural scene. Our emphasis and intent is scene analysis and object identification in the face of 3-D distortions using range sensor data. The range image is segmented into background/nonbackground pixels based on line-by-line processing. Non-background pixels are clustered together by a projection-based technique to determine possible regions of interest. Following extraction of object regions of interest, we can merge the range and other sensor image data to obtain multi-sensor images of the regions in the field of view. Emphasis in this paper is given to the aforementioned tasks; however, for completeness, the remainder of the full processor is briefly described. The use of a hybrid optical/digital processor for such tasks is noted.

An efficient distributed processing scheme has been developed for visual road boundary tracking by 'VaMoRs', a testbed vehicle for autonomous mobility and computer vision. Ongoing work described here is directed to improving the robustness of the road boundary detection process in the presence of shadows, ill-defined edges and other disturbing real world effects. The system structure and the techniques applied for real-time scene analysis are presented along with experimental results. All subfunctions of road boundary detection for vehicle guidance, such as edge extraction, feature aggregation and camera pointing control, are executed in parallel by an onboard multiprocessor system. On the image processing level local oriented edge extraction is performed in multiple 'windows', tighly controlled from a hierarchically higher, modelbased level. The interpretation process involving a geometric road model and the observer's relative position to the road boundaries is capable of coping with ambiguity in measurement data. By using only selected measurements to update the model parameters even high noise levels can be dealt with and misleading edges be rejected.

This paper describes recent research results in the area of autonomous robotics. A multiresolutional (pyramidal) system of world representation is applied to the domain required for Intelligent Mobile Autonomous Systems for outdoor application. Problems of generalization, accuracy, and adequacy of representation are formulated for a system with vision and ultrasonic sensors. A method of nested hierarchical planning is applied for this system of representation. A Pilot with behavioral duality is proven to be efficient for a mobile robot. An execution controller is explored based upon a production system providing near-minimum time operation.

This paper presents a methodology for a concise representation of the 3-D world model for a mobile robot, using range data. The process starts with the segmentation of the scene into "objects" that are given a unique label, based on principles of range continuity. Then the external surface of each object is partitioned into homogeneous surface patches. Contours of surface patches in 3-D space are identified by estimating the normal and curvature associated with each pixel. The resulting surface patches are then classified as planar, convex or concave. Since the world model uses a volumetric representation for the 3-D environment, planar surfaces are represented by thin volumetric polyhedra. Spherical and cylindrical surfaces are extracted and represented by appropriate volumetric primitives. All other surfaces are represented using the boolean union of spherical volumes (as described in a separate paper by the same authors). The result is a general, concise representation of the external 3-D world, which allows for efficient and robust 3-D object recognition.

This paper describes a 3-dimensional vision system that utilizes two views of a scene and textural information in determining distances to objects in a scene. The application of the system lies with determining distances to obstacles in mobile robot guidance. The analysis of a scene is performed in two stages. First, orientations of planar textured surfaces in a scene are estimated using textural information in a single image frame. This step determines the relative positions of obstacles located on textured surfaces. In the second step, exact positions of the obstacles are determined using two views of a scene.

A simple, low cost, infra-red rangefinder has been developed for investigation of autonomous robot cart navigation in factories and similar environments. A 2mW 0.84 LED source (not a laser) is 100% amplitude modulated at 5MHz and used to form a collimated 1" diameter transmit beam that is unconditionally eye-safe. Returning scattered radiation is focussed by a 4" diameter coaxial Fresnel lens onto a p-i-n silicon photodiode. Range is determined from the phase shift between the 5MHz modulation on the transmitted and received signals. Use of a rotating mirror provides 360° polar coordinate coverage of both distance and reflectance out to "20 ft. around the vehicle. Both radial and angular resolution correspond to ~l inch at a ranging distance of 5 ft., with an overall bandwidth of ~lKHz. The ranging resolution at 20 ft. is '2.5 inches, which is close to the theoretical limit possible for the radiated power, bandwidth, optics and receiver employed. The system is capable of reading wall bar codes "on the fly" and is in addition capable of simultaneously ranging and acting as a wideband optical communication receiver. Total parts cost for the optical transmitter, Fresnel lens, receiver and all the electronics is <$200. The remaining major parts, consisting of the rotating mirror, ring mounting, motor and incremental encoder, cost <$500.

Vision systems for mobile robots or autonomous vehicles navigating in an unknown terrain environment must provide a rapid and accurate method of segmenting the scene ahead into regions of pathway and background. A major distinguishing feature between the pathway and background is the three dimensional texture of these two regions. Typical methods of textural image segmentation are very computationally intensive, often lack the required robustness, and are incapable of sensing the three dimensional texture of various regions of the scene. A method is presented where scanned laser projected lines of structured light, viewed by a stereoscopically located single video camera, resulted in an image in which the three dimensional characteristics of the scene were represented by the discontinuity of the projected lines. This image was conducive to processing with simple regional operators to classify regions as pathway or background. Design of some operators and application methods, and demonstration on sample images are presented. This method provides rapid and robust scene segmentation capability that has been implemented on a microcomputer in near real time, and should result in higher speed and more reliable robotic or autonomous navigation in unstructured environments.

We introduce methods for reconstruction of three dimensional surfaces. They are based on moving the light source relative to an object whose shape is to be determined. Using a camera which is placed above the object and a moving light source, the shadows cast by the object at each angle are recorded and analyzed. In one method, the light source is rotated in a horizontal plane; this method is a simple way to get segmentation between different surfaces. The second method uses a light source which is rotated in a vertical plane. For each section of the object a shadow diagram (Shadowgram) is formed and analyzed to get the third dimension. The Shadowgram has some features which make the reconstruction very simple. By looking at some curves of the Shadowgram, invisible surfaces can be partially reconstructed. The above methods can be combined to achieve simple solutions for problems in the robotics area e.g., the bin-picking problem. A set of experimental results demonstrates the robustness and usefulness of the methods.

More than a decade has passed since the Viking landings explored the surface of Mars. Now, after this significant hiatus, momentum is building for a return to the Martian surface. NASA is studying a Mars Rover and Sample Return mission to perform in-situ analysis and collect surface samples of great geological diversity. By their very nature, sites with high scientific interest are not easily accessible from each other or from a safe landing area (large, smooth regions). Consequently, success of the science mission is critically dependent on technological capabilities in the areas of rover mobility, hazard avoidance and navigation. This paper presents a preliminary analysis of the key technical issues and presents representative design approaches.

A Mars rover must be able to sense its local environment with sufficient resolution and accuracy to avoid local obstacles and hazards while moving a significant distance each day. Power efficiency and reliability are extremely important considerations, making stereo correlation an attractive method of range sensing compared to laser scanning, if the computational load and correspondence errors can be handled. Techniques for treatment of these problems, including the use of more than two cameras to reduce correspondence errors and possibly to limit the computational burden of stereo processing, have been tested at JPL. Once a reliable range map is obtained, it must be transformed to a plan view and compared to a stored terrain database, in order to refine the estimated position of the rover and to improve the database. The slope and roughness of each terrain region are computed, which form the basis for a traversability map allowing local path planning. Ongoing research and field testing of such a system is described.

The Martian soil type, slope, and rock size distributions and moisture content are similar to those of the American West. Notable exceptions include significantly reduced gravity, wind pressure, water content and temperature. Previously published pictures of Mars Rover designs proffer concepts unlike those of conventional earth-based vehicles. Novel mobility systems, including jet planes, bag wheel axles, walking machines, articulated wheeled and track modules, balloons and dirigibles have been suggested. The locomotion systems of these vehicles are compared on the basis of energy utilization (mileage) and mobility. A novel single track crawler is described. Steering of the single track crawler is accomplished with a laterally flexible track and cable system.

Data from the Viking missions paint a terrain of volcanic plains, complex lava flow, channel beds, and escarpments. This extreme variety of terrain puts a stringent requirement on the Mars rover design. The rover's versatility will be reflected by its performance over troughs, through rock fields, over duricrust, and about craters. The terrain to be traversed will consist of boulders, and surfaces, loose soil, and partially buried rocks. Wheeled, tracked, or legged locomotion are some of the options to be examined for the rover's primary mode of mobility. This paper will discuss the relationship of the vehicle to this terrain. In particular, what should the feet of a walking vehicle look like? Should the wheels or tracks be ribbed or channeled? As a result, other possible configurations will be looked at with respect to sinkage, slip, obstacle crossing, slope performance, and drawbar pull/weight ratio. The preliminary analysis will be key to design of the Mars mobility system and set a guideline for future concepts.

This paper presents the vision system and the algorithms used to process range data for terrain following of a legged walking machine. The vision system consists of a laser range-finder, a vision computer, a terrain elevation map, and a guidance computer. The range data generated by the laser range-finder are processed and converted into a 3-D representation by the vision computer in real time, and then the elevation information along with the time data are stored in the terrain elevation map. With the real-time elevation information in the map, the guidance computer can select the best footholds for the walking machine in order to maneuver over rough terrain.

We propose a method for the recovery of motion parameters of a rigid object moving through environment with constant but arbitrary linear and angular velocities. The method uses temporal information from a sequence of images such as those taken by a mobile robot. Spatial information contained in the images is also used. The temporal sequence, combined with the assumption of constant velocities, provides powerful constraints for the motion trajectory of rigid objects. We derive a closed form solution for the rigid object trajectory by integrating the differential equations describing the motion of a point on the tracked object. The integrated equations are non-linear only in angular velocity and are linear in all other motion parameters. These equations allow the use of a simple least-square error minimization criterion during an iterative search for the motion parameters. Experimental results demonstrate the power of our method in fast and reliable convergence.

In this paper, a novel method of recovering 3D structure and motion parameters using only the apparent motion in the 2D image plane (Optical Flow) is applied to an extended sequence of (real) images. The details of the method are briefly reviewed, and a controlled laboratory image sequence used to establish its performance in practice. The magnitude and direction of the camera's linear and angular velocities are determined, together with an estimate of the depth to any point in the image. These are compared with the known (measured) values for these quantities. The effects of noisy data are examined, and it is concluded that the method is robust in the presence of image noise. Results from a longer "real" sequence are also presented-specifically of a camera flying over realistic terrain. The method is stable over time and is relatively efficient. Possible applications of this method to passive navigation and the guidance of a robot arm are also considered.

It is stated in the recent published literature that the recovery of structure and motion of moving objects using line correspondences in two frames is ambiguous, if not impossible. In this paper, a method is presented which uses correspondences of five lines in two views of a moving object to calculate its structure, axis of rotation, amount of rotation, and translation vector. The only assumption made is that the object is rigid; and the only information used is the directions of the matched line segments. First, the structure of the moving object (i.e., the relative position of the line segments in space) is determined by solving a set of polynomial equations. These equations are based on the following property of a rigid object: the angle and distance between any two lines on the object do not change from frame to frame. Once the relative positions of the lines in space are recovered, the axis of rotation, the amount of rotation, and the translation vector are easily recovered (in that order) by solving a set of linear equations for each parameter.

Using a discrete distance transform one can quickly build a map of the distance from a goal to every point in a digital map. Using this map, one can easily solve the shortest path problem from any point by simply following the gradient of the distance map. This technique can be used in any number of dimensions and can incorporate obstacles of arbitrary shape (represented in the digital map) including pseudo-obstacles caused by unattainable configurations of a robotic system. This paper further extends the usefulness of the digital distance transform technique by providing an efficient means for dealing with objects which undergo motion. In particular, an algorithm is presented that allows one to update only those portions of the distance map that will potentially change as an object moves. The technique is based on an analysis of the distance transform as a problem in wave propagation. The regions that must be checked for possible update when an object moves are those that are in its "shadow", or in the shadow of objects that are partially in the shadow of the moving object. The technique can handle multiple goals, and multiple objects moving and interacting in an arbitrary fashion.

Cost appraisal is one of the most important elements in route planning, especially when dynamic programming is employed in the search technique. Different missions require different routes; for example, a mission may require a minimum-length path, a terrain-constrained path, a tactical path, a threat-avoidance path, a fuel-constrained path, or a combination of such paths. An optimal path is the path whose total cost is appraised to be the minimum. One difficulty in cost appraisal is in identifying common as well as mission-specific parameters. Another difficulty is in ranking the weighting coefficients, or equivalently the importance factors. Many papers can be found on finding an optimal path for some cost function. However, there has been much less published analysis discussing the detailed definitions of cost-measuring functions other than the familiar Euclidean metrics. Known methods of cost appraisal tend to be either exclusively heuristic or numerical. In this paper we present a new formulation of cost appraisal in which intertwined employment of both approaches is evident. An application example is given.

The paper introduces some improvements over an heuristic approach previously proposed by the authors to tackle the problem of finding a collision-free path for a rigid body amidst a set of fixed obstacles in a 2D environment. While the former approach was based on the guiding of the cartesian degrees of freedom of the mobile body during the search for a path in configuration space, the new one resorts also to the guiding of the rotational degree of freedom of the mobile body. This is a natural extension of the previous approach which, while preserving its efficiency, leads to paths with better clearance properties. To shorten of the resulting solution paths and to improve their smoothness, relaxation techniques are used and shown to convey interesting possibilities as a path planning postprocessing step.

An algorithm is presented for planning a 2-D collision-free path for a mobile robot in an unstructured work environment. The algorithm assumes the existence of a pixel map of all or part of the environment, where each pixel is either on (implying blocked) or off (implying clear). The goal is to compute a "reasonable path" between two points in a minimal amount of time, and this is achieved through a "compressed" representation of the pixel map using a modified quadtree data structure and a procedure for smoothing the resulting path. The algorithm has been coded in the C programming language to facilitate portability, and the results of tests made on "realistic" indoor environments are presented. A discussion on how the environmental maps were obtained is also included.

A robust, non-linear filter with parametric memory is applied to the case of a mobile robot with inaccurate dead reckoning. The robot occasionally receives ambiguous information from known landmarks. The ability of the filter to fuse data in space and time is demonstrated and parametric navigation performance is presented.

A computer system which can find the pathway in the robot's sensor field of view with only one camera has been developed. The low level extraction of linear segmentations is performed by edge following algorithm with the local gray level and gradient as heuristics. The symbolic lists which describe the optical and geometric properties of linear segmentations are generated as the output of the low level and submit to the high level. In high level stage a reasoning mechanism whose knowledge base consists of the rules modeling three basic types of pathway segments (non-branch, branch, and end) works on the symbolic lists to make the judgement of which type to be happened in the sensor area. Several parameters such as the exact positions of the pathway can be calculated as the signals for navigation control system. The experiments showed that the system behaved specially well for the existence of noise. shadow. and irregularities of environments.

Traditional hierarchical robot control systems, although well suited for manufacturing applications, appear to be inefficient for innovative applications, such as mobile robots. The research we present aims to the development of a new architecture, designed to overcome actual limitations. The control system was named BARCS (Behavioral Architecture Robot Control System). It is composed of several modules, that exchange information through a blackboard. The original point is that the functions of the modules were selected according to a behavioral rather than a functional decomposition model. Therefore, the system includes, among other, purpose, strategy, movement, sensor handling and safety modules. Both the hardware structure and the logical decomposition allow a great freedom in the design of each module and of the connections between modules, that have to be as flexible and efficient as possible. In order to obtain an "intelligent" behavior, a mixture of traditional programming, artificial intelligence techniques and fuzzy logic are used, according to the needs of each moddle. The approach is particularly interesting because the robot can be quite easily "specialized", i.e. it can be given behaviors and problem solving strategies that suit some applications better than other. Another interesting aspect of the proposed architecture is that sensor information handling and fusion can be dynamically tailored to the robot's situation, thus eliminating all time-consuming useless processing.

Mobile mechanical devices working with inspection and maintenance of pipes in electrical power plants might become useful to E.D.F. A specific field in mobile robotics, with its proper know-how and complications, is being developed. Simulations of different aspects of robot conception are presented : - environment modelling, - kinetics : modelling which can include closed chains, - navigation and obstacle avoidance : sensor choice and positionning, strategy control.

Navigation and visual guidance are key topics in the design of a mobile robot. Omnidirectional vision using a very wide angle or fisheye lens provides a hemispherical view at a single instant that permits target location without mechanical scanning. The inherent image distortion with this view and the numerical errors accumulated from vision components can be corrected to provide accurate position determination for navigation and path control. The purpose of this paper is to present the experimental results and analyses of the imaging characteristics of the omnivision system including the design of robot-oriented experiments and the calibration of raw results. Errors less than one picture element on each axis were observed by testing the accuracy and repeatability of the experimental setup and the alignment between the robot and the sensor. Similar results were obtained for four different locations using corrected results of the linearity test between zenith angle and image location. Angular error of less than one degree and radial error of less than one Y picture element were observed at moderate relative speed. The significance of this work is that the experimental information and the test of coordinated operation of the equipment provide a greater understanding of the dynamic omnivision system characteristics, as well as insight into the evaluation and improvement of the prototype sensor for a mobile robot. Also, the calibration of the sensor is important, since the results provide a cornerstone for future developments. This sensor system is currently being developed for a robot lawn mower.

The large nursing home market has several natural characteristics which make it a good applications area for robotics. The environment is already robot accessible and the work functions require large quantities of low skilled services on a daily basis. In the near future, a commercial opportunity for the practical application of robots is emerging in the delivery of housekeeping services in the nursing home environment. The robot systems will assist in food tray delivery, material handling, and security, and will perform activities such as changing a resident's table side drinking water twice a day, and taking out the trash. The housekeeping work functions will generate cost savings of approximately $22,000 per year, at a cost of $6,000 per year. Technical system challenges center around the artificial intelligence required for the robot to map its own location within the facility, to find objects, and to avoid obstacles, and the development of an energy efficient mechanical lifting system. The long engineering and licensing cycles (7 to 12 years) required to bring this type of product to market make it difficult to raise capital for such a venture.

We describe a method of autonomous programming for mobile robots. A set of procedures are defined for a small class of features. The robot searches for these features, and determines their locations. The features and their locations comprise a world model. Errors in dead reckoning arise during navigation through the environment. These errors result in uncertainty in the feature locations. A Prolog-based world model is used to represent the feature locations, thereby determining the robot's path through its environment.

An energy boost system based on a flywheel has been designed to supply the surge power needs of mobile robots for operating equipment like transmitters, drills, manipulator arms, mobility augmenters, and etc. This flywheel increases the average power available from a battery, fuel cell, generator, RPG or solar array by one or more orders of magnitude for short periods. Flywheels can be charged and discharged for thousands of battery lifetimes. Flywheels can deliver more than ten times the power per unit weight of batteries. The electromechanical details of a reliable, energy efficient and (relatively) low cost flywheel are described. This flywheel is the combination of a highly efficient brushless motor and a laminated steel rotor operating in an hermetically sealed container with only electrical input and output. This design approach overcomes the inefficiencies generally associated with mechanically geared devices. Electrical round trip efficiency is 94% under optimum operating conditions.

Initial investigations into two different approaches for applying autonomous ground vehicle technology to the vehicle convoying application are described. A minimal capability system that would maintain desired speed and vehicle spacing while a human driver provided steering control could improve convoy performance and provide positive control at night and in inclement weather, but would not reduce driver manpower requirements. Such a system could be implemented in a modular and relatively low cost manner. A more capable system would eliminate the human driver in following vehicles and reduce manpower requirements for the transportation of supplies. This technology could also be used to aid in the deployment of teleoperated vehicles in a battlefield environment. The needs, requirements, and several proposed solutions for such an Attachable Robotic Convoy Capability (ARCC) system will be discussed. Included are discussions of sensors, communications, computers, control systems and safety issues. This advanced robotic convoy system will provide a much greater capability, but will be more difficult and expensive to implement.

In the rapid pursuit of automation, it is sometimes overlooked that an elaborate human-machine interplay is still necessary, despite the fact that a fully automated system, by definition, would not require a human interface. In the future, real-time sensing, intelligent processing, and dextrous manipulation will become more viable, but until then it is necessary to use humans for many critical processes. It is not obvious, however, how automated subsystems could account for human intervention, especially if a philosophy of "pure" automation dominates the design. Teleoperation, by contrast, emphasizes the creation of hardware pathways (e.g., hand-controllers, exoskeletons) to quickly communicate low-level control data to various mechanisms, while providing sensory feedback in a format suitable for human consumption (e.g., stereo displays, force reflection), leaving the "intelligence" to the human. These differences in design strategy, both hardware and software, make it difficult to tie automation and teleoperation together, while allowing for graceful transitions at the appropriate times. In no area of artifical intelligence is this problem more evident than in computer vision. Teleoperation typically uses video displays (monochrome/color, monoscopic/ stereo) with contrast enhancement and gain control without any digital processing of the images. However, increases in system performance such as automatic collision avoidance, path finding, and object recognition depend on computer vision techniques. Basically, computer vision relies on the digital processing of the images to extract low-level primitives such as boundaries and regions that are used in higher-level processes for object recognition and positions. Real-time processing of complex environments is currently unattainable, but there are many aspects of the processing that are useful for situation assessment, provided it is understood the human can assist in the more time-consuming steps.