Traditionally, path planning for field robotic systems is performed in Cartesian space: sensor readings are transformed into terrain costs in a (Cartesian) costmap, and a path to the goal is planned in that map. In this paper, we propose a new approach: planning a path for the robot in the image-space of an on-board camera. We apply a learned color- to-cost mapping to transform a raw image into a cost-image, which then undergoes a pseudo-configuration-space transform. We search in the resulting cost-image for a path to the projected goal point in the image. One benefit of our approach is the ability to react to obstacles at ranges well beyond our 3D sensor range - independent testing has confirmed our system has effectively reacted to obstacles at a range of 93 m while our stereo sensor provides reliable data only up to 5 m away. We describe the details of our technique and the results from testing under the DARPA LAGR and UPI programs.

We describe two new sampling strategies for Rao-Blackwellized particle filtering SLAM. The strategies, called fixed-lag roughening and the block proposal distribution, both exploit "future" information, when it becomes available, to improve the filter's estimation for previous time steps. Fixed-lag roughening perturbs trajectory samples over a fixed lag time according to a Markov chain-Monte Carlo kernel. The block proposal distribution directly samples poses over a fixed lag from their fully joint distribution conditioned on all the available data. Our experimental results indicate that the proposed strategies, especially the block proposal, yield significant improvements in filter consistency and a reduction in particle degeneracies compared to standard sampling techniques such as the improved proposal distribution of FastSLAM 2.

Driving in unknown natural outdoor terrain is a challenge for autonomous ground vehicles. It can be difficult for a robot to discern obstacles and other hazards in its environment, and characteristics of this high cost terrain may change from one environment to another, or even with different lighting conditions. One successful approach to this problem is for a robot to learn from a demonstration by a human operator. In this paper, we describe an approach to calculating terrain costs from Bayesian estimates using feature vectors measured during a short teleoperated training run in similar terrain and conditions. We describe the theory, its implementation on two different robotic systems, and results of several independently conducted field tests.

Most work on the simultaneous localization and mapping (SLAM) problem assumes the frequent availability of dense information about the environment such as that provided by a laser rangefinder. However, for implementing SLAM in consumer-oriented products such as toys or cleaning robots, it is infeasible to use expensive sensing. In this work we examine the SLAM problem for robots with very sparse sensing that provides too little data to extract features of the environment from a single scan. We modify SLAM to group several scans taken as the robot moves into multiscans, achieving higher data density in exchange for greater measurement uncertainty due to odometry error. We formulate a full system model for this approach, and then introduce simplifications that enable efficient implementation using a Rao-Blackwellized particle filter. Finally, we describe simple algorithms for feature extraction and data association of line and line segment features from multiscans, and then present experimental results using real data from several environments

In order to create consistent maps of unknown environments, a robot must be able to recognize when it has returned to a previously visited place. In this paper, we introduce an evidential approach to the loop-closing problem for topological maps, based on the Dempster-Shafer theory of evidence. In our approach, the robot makes a hypothesis whenever it may have revisited a place. It then attempts to verify hypotheses by continuing to traverse the environment, gathering evidence that supports (or refutes) the hypotheses. We describe methods for managing belief about multiple loop-closing hypotheses, and for determining a belief assignment given a piece of evidence. We also discuss methods for reducing the false alarm rate of our loop-closing algorithm, and provide simulated and real-world experimental results that verify the effectiveness of our approach.

This paper explores a novel mode of mobile manipulation using a car-like robot whose chassis does not extend above the height of its wheels. It transports an object that rests on top of the wheels, so as the robot moves forwards, the object moves forward relative to the robot. The robot must maneuver so that the object stays on top of the robot, or a second robot must be in position to receive the robot as it rolls off the first one. We describe the mechanics of this system, its basic maneuvers and manipulation techniques, present a motion planner for generating plans to transport an object, and give results of our experimental implementation of this system.

Robotic coverage is the problem of moving a sensor or actuator over all points in given region. Ultimately, we want a coverage path that minimizes some cost such as time. We take the approach of decomposing the coverage region into subregions, selecting a sequence of those subregions, and then generating a path that covers each subregion in turn. We focus on generating decompositions based upon the planar line sweep. After a general overview of the coverage problem, we describe how our assumptions lead to the optimality criterion of minimizing the sum of subregion altitudes (which are measured relative to the sweep direction assigned to that subregion). For a line-sweep decomposition, the sweep direction is the same for all subregions. We describe how to find the optimal sweep direction for convex polygonal worlds. We then introduce the minimal sum of altitudes (MSA) decomposition in which we may assign a different sweep direction to each subregion. This decomposition is better for generating an optimal coverage path. We describe a method based on multiple line sweeps and dynamic programming to generate the MSA decomposition.

Manipulation is a key capability for mobile robots to perform a greater variety of useful tasks. We consider mobile robots equipped with low degree-of-freedom "palm manipulators." The advantage of using this sort of nonprehensile (i.e. nongrasping) manipulator is its simplicity and versatility. Although in general multiple robots may be needed, the initial work described in the paper considers only a single robot. We formulate the mechanical analysis for manipulating objects and describe a planner for a single robot with a palm manipulator. We conclude with details of our experimental implementation and preliminary results.

Describes the use of tapping actuators in a "micropositioning cell." Tapping actuators are fixed (both in position and orientation) about the perimeter of the cell. A part in the cell can be precisely positioned by firing a sequence of tapping actuators (sensing the position of the object after each tap for feedback control). Such a system could be used for micropositioning tasks or in parts feeding applications. We show the controllability of this system for circular parts, describe an algorithm to position these objects, and present the results of simulated experiments. We conclude with some future directions and address the issues in extending this work to more efficient algorithms that work on a broader class of parts.

In this paper, we present the results of our experimental effort in one form of impulsive manipulation: tapping. Our previous work studied the mechanics of tapping a planar object which then slides on a support surface, coming to rest due to friction. This work addresses the practical issues in creating a system which uses this mode of manipulation. We begin with the design of tapping devices-end effecters designed to deliver an impulse to an object, and report some of the issues we have found to be important in their design. Our next step was to perform single-tap experiments in order to fit and evaluate the models of impact and sliding. These experiments have shown that objects rotate less than predicted; we have found that the addition of a scaling factor for the torque due to friction enables the models to predict object motion reasonably well. In order to do positioning experiments, we developed a number of planning methods (or feedback control strategies) to compensate for errors in modeling, parameters, and actuation. These planning methods were successfully used to demonstrate a positioning task. We also have experimentally demonstrated that tapping can be used to position an object more precisely than the manipulator can position the tapping device. We offer some sensitivity analysis in support of this result.

This paper explores a sensorless technique for orienting planar parts. We follow an approach described in the ours earlier papers (1995), called one joint over conveyor (1JOC), which can perform planar manipulation using a single controlled joint in combination with a constant-velocity conveyor. Our previous work demonstrated that the IJOC approach can orient and feed planar polygonal parts, given a singulated part in a known initial location. This paper shows that a variation called the sensorless IJOC can orient and feed polygonal parts up to symmetries in the underlying mechanics, without knowing the initial location and without sensors.

Planning the placement of stereo cameras on a mobile robot is often a balancing act between the quality of stereo data, the reliability of stereo matching, and ensuring that the cameras see enough of the robot's environment for navigation. In this paper, we present a nonlinear programming formulation to determine the optimal parameters for two stereo pairs of cameras and a camera mast for a mobile robot. We have applied this method to the design of a stereo camera mast for Nomad, a prototype rover currently under construction at Carnegie Mellon University, and present the results of designing that mast.

Vibratory manipulation is any mode of manipulation involving repeated impacts due to a striker which follows some periodic motion. In this paper, we study vibratory manipulation in the context of tapping planar objects which slide on a fixed support surface. We are interested in the behaviors an object exhibits under such excitation. There are two distinct types of tapping that can result: continuous tapping, in which the object is always in motion, and intermittent tapping, in which the object comes to rest between taps. We first examine vibratory manipulation in one dimension, adapting results from related work to find conditions for stable periodic motion. The general two dimensional case is closely related to our previous work in impulsive manipulation which examined the mechanics of a sliding rotating object. In fact, vibratory manipulation is an approximation to the limiting cases of impulsive manipulation. We develop the limiting cases for intermittent and continuous tapping for rotationally symmetric objects and conclude with some examples.

Examines a little-studied method of manipulation-manipulation by striking an object and letting it slide. There are two parts to this problem: The inverse sliding problem, determining the velocities required to send an object to a desired configuration, and the impact problem, determining how to strike the object in order to achieve those velocities. The authors present a solution to these two problems for the class of rotationally symmetric objects and conclude with some observations about this method of manipulation.