Sabastian Thrun. A bayesian approach to landmark discovery and active perception in mobile robot navigation. Technical Report, Carnegie Mellon University, School of Computer Science, CMU-CS-96-122, May 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....back to a second verification procedure, namely using the speech module to ask a human. We assume that verification step gives complete and correct information about the robot s actual location; other researchers are focussing on the open problem of sensor reliability [Hughes ; Ranganathan, 1994, Thrun, 1996] If ROGUE detects that in fact the robot is not at the correct goal location, ROGUE updates the navigation module with the new information and re attempts to navigate to the desired location. If the robot is still not at the correct location after a constant number of tries (three in our ....

Thrun, S. (1996). A Bayesian approach to landmark discovery and active perception for mobile robot navigation. Technical Report CMU-CS-96-122, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA.

....of landmarks would overcome the acquisition rate problem by not requiring the entire visual field to remain constant, therefore maximising the area in which one set of landmarks can be used for localisation. Obviously the success of localisation using landmarks depends on the choice of landmarks [ Thrun, 1996 ] Bianco and Zelinsky, 1999 ] describe a monocular system where landmarks are chosen on the basis of their reliability. To be selected, landmarks must display uniqueness in the immediate surroundings and the ability to remain reliable as the robot moves through the environment. To this end the ....

S. Thrun. A bayesian approach to landmark discovery and active perception in mobile robot navigation. Technical report, School of Computer Science Carnegie Mellon University, 1996.

....sequence approach can be thought of as using one big landmark for the navigation tasks, this avoids the problems of small templates but suffers from high image acquisition rates and high storage cost. Obviously the success of navigation using small landmarks depends on the choice of landmarks [ Thrun, 1996 ] Bianco and Zelinsky, 1999 ] describe a monocular system where landmarks are chosen on the basis of their reliability. To be selected, landmarks must display uniqueness in the immediate surroundings and the ability to remain reliable as the robot moves through the environent. To this end the ....

S. Thrun. A bayesian approach to landmark discovery and active perception in mobile robot navigation. Technical report, School of Computer Science Carnegie Mellon University, 1996.

....environment for localization, may not easily be adapted for outdoor environments which require other image features for navigation. This is because most current approaches require a human to pick specific landmarks by which to navigate, thus making the localization system application specific [10], often requiring a complete redesign for each new environment. In this paper we propose an approach to vision based mobile robot localization which is similar to that proposed in [10] We make use of an on board camera and a temporary calibration setup which continuously reports the robot s ....

....a human to pick specific landmarks by which to navigate, thus making the localization system application specific [10] often requiring a complete redesign for each new environment. In this paper we propose an approach to vision based mobile robot localization which is similar to that proposed in [10]. We make use of an on board camera and a temporary calibration setup which continuously reports the robot s position, to learn the mapping between the on board camera s image and its position. During this calibration phase, the robot is directed to move through its environment and its position ....

[Article contains additional citation context not shown here]

Sabastian Thrun. A bayesian approach to landmark discovery and active perception in mobile robot navigation. Technical Report, Carnegie Mellon University, School of Computer Science, CMU-CS-96-122, May 1996.

....To be truly autonomous, the robot needs to be able to use accumulated experience and feedback about its performance to improve its behaviour. It needs to learn. Learning has been applied to robotics problems in a variety of manners. Common applications include map learning and localization (e.g. [12, 13, 25]) or learning operational parameters for better actuator control (e.g. 2, 4, 19] Instead of improving low level actuator control, our work focusses at the planning stages of the system. A few other researchers have explored this area as well, learning costs of actions, or their applicability ....

S. Thrun. A Bayesian approach to landmark discovery and active perception for mobile robot navigation. Technical Report CMU-CS-96-122, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, 1996.

....Localization techniques allow the robot to place itself within the map, and therefore in its environment. Simple localization techniques, such as dead reckoning, accumulate odometry error quickly. More advanced localization techniques typically depend on recognizing and in some cases learning [6] individual landmarks in the environment. Researchers have looked at localization techniques that allow the use of the same maps being used for navigation [7] However, in these cases, changes to the environ ment can have a detrimental effect on the ability to localize or navigate. Continuous ....

Thrun, S., "A Bayesian approach to landmark discovery and active perception in mobile robot navigation," Technical Report CMU-CS-96-122, Carnegie Mellon University, Pittsburgh, PA, 1996.

....back to a second verification procedure, namely using the speech module to ask a human. We assume that verification step gives complete and correct information about the robot s actual location; other researchers are focussing on the open problem of sensor reliability [Hughes Ranganathan, 1994, Thrun, 1996] If Rogue detects that in fact the robot is not at the correct goal location, Rogue updates the navigation module with the new information and re attempts to navigate to the desired location. If the robot is still not at the correct location after a constant number of tries (three in our ....

Thrun, S. (1996). A Bayesian approach to landmark discovery and active perception for mobile robot navigation. Technical Report CMU-CS-96-122, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA.

.... measurement than that associated with the optimal straight line path [42] In artificial landmark recognition, uniquely determinable and specially designed markers or objects are placed in various a priori known locations in the environment for the sole purpose of enabling robot navigation [75]. The advantage of such a method is that the robot location can be uniquely determined even in adverse conditions. Triangulation can also be performed (when three or more landmarks are visible) to get exact positions. The biggest disadvantage of this approach is that it can be very inaccurate when ....

....distinctive features that are an inherent part of the environment. The environment must be known in advance but does not have to be engineered in any manner. Some added disadvantages include increased computation time due to more complex object recognition tasks and decreased reliability [29] [75]. This is especially true if one uses natural landmarks in an outdoor environment since the time required to recognize non geometric and arbitrarily shaped landmarks (typically present in outdoor environments) is much higher than that required for recognizing linear landmarks (peg. doors and ....

[Article contains additional citation context not shown here]

S, Thrun. A Bayesian Approach to Landmark Discovery and Active Perception in Mobile Robot Navigation. Technical Report CMU-CS-96-122, Carnegie Mellon University, Pittsburgh, PA, 1996.

.... be transformed to (9) Substituting (9) into (8) yields a recursive rule for the computation of all ff with boundary condition (7) which uses the data d, the model m, in conjunction with the motion model P ( ju; and the perceptual model P (jo; m) See [26] for a more detailed derivation. 3.1.2 Computation of the fi Values The computation of fi is completely analogous, but backwards in time. The initial fi , which expresses the probability that the robot s final position is is uniformly distributed (fi does not depend on data) All ....

S. Thrun. A bayesian approach to landmark discovery and active perception for mobile robot navigation. Technical Report CMU-CS-96-122, Carnegie Mellon University, School of Computer Science, Pittsburgh, PA 15213, April 1996.

....to a particular environment and to its sensors, making it more widely applicable than methods that rely on static, built in landmarks. 2 Approach This section presents our learning algorithm for the automatic selection of landmarks. Due to brevity, the algorithm is only sketched here; see [16] for the complete description and mathematical derivation. The input to our algorithm is a set of sensor snapshots along with the position at which they were taken. These data is used to establish the correspondence between world coordinates and sensor values and to select appropriate landmarks. ....

....at random, then use the data to construct P (sj) compute E post , and then pick those weights that minimize E post . However, random sampling is of course hopelessly inefficient. Our learning approach utilizes the fact that E post is differentiable in the weights of the neural networks (see [16] for an exact derivation of the derivatives of E post ) To train the networks, each weight w ij is adjusted in proportion to the negative gradient Gammar w ij E post (just like weights and biases are adjusted in Back Propagation) Repeated application of this gradient descent update rule leads ....

[Article contains additional citation context not shown here]

S. Thrun. A Bayesian Approach to Landmark Discovery and Active Perception for Mobile Robot Navigation TR CMU-CS96 -122, 1996.

.... global uncertainty, a problem which is also known as the kidnaped robot problem [33] Only a small number of localization methods are capable of localizing a robot under global uncertainty, and all of those require (for obvious reasons) that the robot be equipped with a map of the environment [7,10 12,100]. 5.3 Decomposition and Robot Motion Planning The topological map extraction algorithm extracts a coarse grained representation from high resolution maps. Within the robot motion planning community, such algorithms are usually referred to as cell decomposition methods [92,60] Within Artificial ....

....information over multiple measurements at multiple locations is automatically done in a consistent way. Visual landmarks, which often come to bear in topological approaches, can certainly be incorporated into the current approach to further improve the accuracy of position estimation (see e.g. [55,100]) In fact, sonar sensors can be understood as landmark detectors that indirectly through the grid based map help determine the actual position in the topological map (cf. 95] One of the key empirical results of this research concerns the cost benefit analysis of topological ....

S. Thrun. A bayesian approach to landmark discovery and active perception for mobile robot navigation. Technical Report CMU-CS-96-122, Carnegie Mellon University, School of Computer Science, Pittsburgh, PA 15213, April 1996.

.... ) ff (t Gamma1) d (t Gamma1) 9) Substituting (9) into (8) yields a recursive rule for the computation of all ff (t) with boundary condition (7) which uses the data d, the model m, in conjunction with the motion model P ( 0 ju; and the perceptual model P ( jo; m) See [26] for a more detailed derivation. 3.1.2 Computation of the fi Values The computation of fi (t) is completely analogous, but backwards in time. The initial fi (T ) which expresses the probability that the robot s final position is is uniformly distributed (fi (T ) does not ....

S. Thrun. A bayesian approach to landmark discovery and active perception for mobile robot navigation. Technical Report CMU-CS-96-122, Carnegie Mellon University, School of Computer Science, Pittsburgh, PA 15213, April 1996.

....8) Once the global wall orientation has been estimated, it is used to readjust the robot s orientation based on future sonar measurements. See [46] for more details. ffl Landmarks. Landmarks are used in various approaches to mobile robot localization (see e.g. 3, 25, 32] and references in [44]) We recently have begun to explore mechanisms that enable a robot to select its own landmarks, based on sonar and camera input. The key idea underlying this approach is to train artificial neural networks to recognize landmarks by minimizing the average localization error (assuming that update ....

....the average localization error (assuming that update rule (2) is applied in localization) As a result, our robot successfully discovered a variety of useful visual landmarks, such as doors, wall color, ceiling lights and so on. Details of the algorithm and performance results are surveyed in [44]. This list of sources for estimating l has been developed over the last few years. Some of these methods make strong assumptions on the correctness of the global map (e.g. the maneuverability method) hence cannot be interleaved with map learning. The reader should also notice that the ....

S. Thrun. A bayesian approach to landmark discovery and active perception for mobile robot navigation. Technical Report CMU-CS-96-122, Carnegie Mellon University, School of Computer Science, Pittsburgh, PA 15213, April 1996.

....information over multiple measurements at multiple locations is automatically done in a consistent way. Visual landmarks, which often come to bear in topological approaches, can certainly be incorporated into the current approach, to further improve the accuracy of position estimation (see e.g. [14, 32]) In fact, sonar sensors can be understood as landmark detectors that indirectly through the gridbased map help determine the actual position in the topological map (cf. 30] One of the key empirical results of this research concerns the cost benefit analysis of topological representations. ....

S. Thrun. A bayesian approach to landmark discovery and active perception for mobile robot navigation. Technical Report CMU-CS-96-122, Carnegie Mellon University, School of Computer Science, Pittsburgh, PA 15213, April 1996.

....avoiding collisions with various obstacles, some of which were invisible. RHINO s probabilistic localization methods [2, 3] provided it with reliable position estimates, which was essential for the robustness of the entire approach. Probabilistic representations also played a key role in BaLL [17], an algorithm 1 This work was carried out jointly with Dirk Hahnel, Dirk Schulz, and Wolli Steiner from the Institut fur Informatik of the Universitat Bonn. See http: www.cs.uni bonn.de RHINO tourguide for further information. Figure 1: The robot RHINO, a B21 robot manufactured by Real ....

....where the exact location of obstacles is known. Figure 4 provides examples of probabilities generated by the networks after training. The darker a value in the circular region around the robot, the higher the probability that the corresponding grid cell is occupied. In a second approach, BaLL [17], neural network learning was employed to map high dimensional data into low dimensional spaces, so that this low dimensional data can be used to model P (sj ) In Ball, a Bayesian learning approach is employed that trains neural networks so that the n most useful bits are extracted for the state ....

S. Thrun. A bayesian approach to landmark discovery and active perception for mobile robot navigation. Technical Report CMU-CS-96-122, Carnegie Mellon University, School of Computer Science, Pittsburgh, PA 15213, April 1996.

No context found.

Sabastian Thrun. A bayesian approach to landmark discovery and active perception in mobile robot navigation. Technical Report, Carnegie Mellon University, School of Computer Science, CMU-CS-96-122, May 1996.

No context found.

S. Thrun, \A bayesian approach to landmark discovery and active perception in mobile robot navigation," Tech. Rep. CMU-CS96 -122, School of Computer Science Carnegie Mellon University, Pittsburgh, PA 15213, May 1996.

No context found.

S. Thrun. A bayesian approach to landmark discovery and active perception in mobile robot navigation. Technical report, School of Computer Science Carnegie Mellon University, 1996.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC