## Learning Maps for Indoor Mobile Robot Navigation (1997)

### Cached

### Download Links

Venue: | ARTIFICIAL INTELLIGENCE (ACCEPTED FOR PUBLICATION) |

Citations: | 88 - 12 self |

### BibTeX

@MISC{Thrun97learningmaps,

author = {Sebastian Thrun},

title = {Learning Maps for Indoor Mobile Robot Navigation},

year = {1997}

}

### Years of Citing Articles

### OpenURL

### Abstract

Autonomous robots must be able to learn and maintain models of their environments. Research on mobile robot navigation has produced two major paradigms for mapping indoor environments: grid-based and topological. While grid-based methods produce accurate metric maps, their complexity often prohibits efficient planning and problem solving in large-scale indoor environments. Topological maps, on the other hand, can be used much more efficiently, yet accurate and consistent topological maps are often difficult to learn and maintain in large-scale environments, particularly if momentary sensor data is highly ambiguous. This paper describes an approach that integrates both paradigms: grid-based and topological. Grid-based maps are learned using artificial neural networks and naive Bayesian integration. Topological maps are generated on top of the grid-based maps, by partitioning the latter into coherent regions. By combining both paradigms, the approach presented here gains advantages from both worlds: accuracy/consistency and efficiency. The paper gives results for autonomous exploration, mapping and operation of a mobile robot in populated multi-room environments.

### Citations

7440 |
Probabilistie Reasoning in Intelligent Systems: Networks of Plausible Inference
- Pearl
- 1988
(Show Context)
Citation Context ...ob(x) ! \Gamma1 Here Prob(x) denotes the prior probability for occupancy (which, if set to 0.5, can be omitted in this equation). The derivation of this formula is straightforward and can be found in =-=[19, 21]-=-. Notice that this formula can be used to update occupancy values incrementally. An example map of a competition ring constructed at the 1994 AAAI autonomous robot competition is shown in Figure 4. 2.... |

4143 |
Artificial Intelligence: A Modern Approach, Pearson Education, 2nd edition
- Russell, Norvig
- 1995
(Show Context)
Citation Context ...rid-based map. For every abstract plan generated using the topological map, there exists a corresponding plan in the grid-based map (in other words, the abstraction has the downward solution property =-=[26]-=-). Conversely, every path that can be found in the grid-based map has an abstract representation which is a admissible plan in the topological map (upward solution property). Notice that although cons... |

4141 |
Pattern classification and scene analysis
- Duda, Hart
- 1973
(Show Context)
Citation Context ...d to minimize cross-entropy; we refer the interested reader to page 167 of Mitchell's textbook [67], which demonstrates that networks trained in this way approximate the maximum likelihood hypothesis =-=[30,107]-=-. Figure 3 shows three examples of sonar scans (top row, bird's eye view) along with their neural network interpretation (bottom row). The darker a value in the circular region around the robot, the l... |

2980 |
R.J.: Learning Internal Representations by Error Propagation
- Rumelhart, Hinton, et al.
- 1986
(Show Context)
Citation Context ...er. 2.1 Sensor Interpretation To build metric maps, sensor reading must be "translated" into occupancy values occ x;y for each grid cell hx; yi. The idea here is to train an artificial neura=-=l network [25]-=- using Back-Propagation to map sonar measurements to occupancy values [31]. As shown in Figure 2, the input to the network consists of ffl two values that encode hx; yi in polar coordinates relative t... |

2856 |
Dynamic Programming
- Bellman
- 1957
(Show Context)
Citation Context ...ed to construct many training examples for different x-y coordinates. In our implementation, training examples are generated by a mobile robot simulator. Once trained, the network generates values in =-=[0; 1]-=- that can be interpreted as probability 1 for occupancy. Figure 3 shows three examples of sonar scans (top row, bird's eye view) along with their neural network interpretation (bottom row). The darker... |

2431 |
A new approach to linear filtering and prediction problems, Trans
- Kalman
- 1960
(Show Context)
Citation Context ...]. There are several attempts to integrate localization and mapping. For example, Leonard, Durrant-Whyte, and Cox [62] proposed a method that interleaves localization and mapping using Kalman filters =-=[48]-=- for position tracking. In their experiments, however, only the mapping component of their approach is demonstrated, leaving open the question as to whether these methods work well together in practic... |

2181 |
Robot Motion Planning
- Latombe
- 1991
(Show Context)
Citation Context ...ll is occupied. More specifically, it contains the belief whether or not the center of the robot can be moved to the center of that cell (it represents the configuration space of the robot, see e.g., =-=[16]-=-). Occupancy values are determined based on sensor readings. This section describes the four major components of our approach to building grid-based maps [31]: 1. Interpretation. Sensor readings are m... |

1906 |
Introduction to the Theory of Neural Computation
- Hertz, Krogh, et al.
- 1991
(Show Context)
Citation Context ...Situations such as the 1 It has been shown that, under certain assumption, a neural network trained to predict a binary random variable approaches the probability distribution of this random variable =-=[11, 36]-=- 6 S. Thrun and A. Bucken 4 sensor values(s (t) ) ... angle(x,y) distance(x,y) Prob(occ xy s (t) ) Figure 2: An artificial neural network maps sensor measurements to probabilities of occupancy. one sh... |

1431 |
System identification: Theory for the user
- LJUNG
- 1980
(Show Context)
Citation Context ...). The reader should note that our approach is highly specialized to learning spatial maps, whereas methods for learning FSAs are targeted at different, more general problems of system identification =-=[64]-=-. Thus, while our approach is clearly better suited for learning maps, it lacks the generality of the FSA identification algorithms. 5.6 Contributions The major contribution of the current paper is a ... |

1351 |
Graph Theory with Applications
- Bondy, Murty
- 1976
(Show Context)
Citation Context ... iteration updates the value of all explored grid cells by the value of their best neighbors, plus the costs of moving to this neighbor (just like A* [75] or Dijkstra's famous shortest path algorithm =-=[6]-=-). Cost is here equivalent to the probability Prob(occ x;y ) that a grid cell hx; yi is occupied. The update rule is iterated. When the update converges, each value V x;y measures the cumulative cost ... |

928 | Introduction to reinforcement learning
- Sutton, Barto
- 1998
(Show Context)
Citation Context ...and as such relate to the rich literature on abstraction in AI. The most closely related work on abstraction can be found in the literature on dynamic programming [4,44,83] and reinforcement learning =-=[2,46,97]-=-. In fact, our motion planning algorithm can be viewed as a model-based version of reinforcement learning [24,98,114]; however, for the sake of consistency with the literature we will refer to it as d... |

632 |
Markov Decision Processes
- Puterman
- 1994
(Show Context)
Citation Context ... cell. The cost for traversing a grid cell is determined by its occupancy value. The minimum-cost path is computed using a modified version of value iteration, a popular dynamic programming algorithm =-=[4,44,83]-=-: (i) Initialization. Unexplored grid cells are initialized with 0, explored ones with 1: V x;y /\Gamma 8 ? ! ? : 0; if hx; yi unexplored 1; if hx; yi explored Grid cells are considered explored if th... |

587 |
Principles of Artificial Intelligence
- Nilsson
- 1980
(Show Context)
Citation Context ...i=\Gamma1;0;1 fV x+;y+i + Prob(occ x+;y+i )g Value iteration updates the value of all explored grid cells by the value of their best neighbors, plus the costs of moving to this neighbor (just like A* =-=[20]-=-). Cost is here equivalent to the probability Prob(occ x;y ) that a grid cell hx; yi is occupied. The update rule is iterated. When the update converges, each value V x;y measures the cumulative cost ... |

562 |
Dynamic Programming and Markov Processes
- Howard
- 1960
(Show Context)
Citation Context ... cell; The cost for traversing a grid cell is determined by its occupancy value. The minimum-cost path is computed using a modified version of value iteration, a popular dynamic programming algorithm =-=[1, 13]-=-: 1. Initialization. Unexplored grid cells are initialized with 0, explored ones with 1: V x;y /\Gamma ( 0; if hx; yi unexplored 1; if hx; yi explored Grid cells are considered explored if their occup... |

558 | Learning to act using real-time dynamic programming
- Barto, Bradtke, et al.
- 1995
(Show Context)
Citation Context ...and as such relate to the rich literature on abstraction in AI. The most closely related work on abstraction can be found in the literature on dynamic programming [4,44,83] and reinforcement learning =-=[2,46,97]-=-. In fact, our motion planning algorithm can be viewed as a model-based version of reinforcement learning [24,98,114]; however, for the sake of consistency with the literature we will refer to it as d... |

519 |
An analysis of timedependent planning
- Dean, Boddy
- 1988
(Show Context)
Citation Context ...ccupied. The update rule is iterated. When the update converges, each value V x;y measures the cumulative cost for moving to the nearest unexplored cell. However, control can be generated at any time =-=[5]-=-, long before value iteration converges. 3. Determine motion direction. To determine where to explore next, the robot generates a minimum-cost path to the unexplored. This is done by steepest descent ... |

509 |
The complexity of robot motion planning
- Canny
- 1987
(Show Context)
Citation Context ... also addressed issues of efficiency. A key difficulty arises from the observation that robot motion planning, in its general definition, is worst-case exponential in the number of degrees of freedom =-=[13,84]-=-. Due to the strong focus on consistency, O()-type complexity, worst case analysis and robots with many degrees of freedom, existing cell decomposition methods usually decompose the free-space in odd ... |

429 | A robot exploration and mapping strategy based on a semantic hierarchy of spatial representations
- Kuipers, Byun
- 1991
(Show Context)
Citation Context ...onments by evenly-spaced grids. Each grid cell may, for example, indicate the presence of an obstacle in the corresponding region of the environment. Topological approaches, such a those described in =-=[7, 14, 15, 17, 22, 35]-=-, represent robot environments by graphs. Nodes in such graphs correspond to distinct situations, places, or landmarks (such as doorways). They are connected by arcs if there exists a direct path betw... |

429 |
High resolution maps from wide angle sonar
- Moravec, Elfes
- 1985
(Show Context)
Citation Context ... segments and polyhedral objects which are matched to a geometric map. Moravec and Elfes, who pioneered the development of occupancy grids, were also the first to use occupancy grids for localization =-=[31,71]-=-. Just like the approach presented here, they proposed building local maps from single sonar scans and matching them to a previously learned (or hand-supplied) global map to identify errors in odometr... |

383 | Practical issues in temporal difference learning
- Tesauro
- 1992
(Show Context)
Citation Context ... found in the literature on dynamic programming [4,44,83] and reinforcement learning [2,46,97]. In fact, our motion planning algorithm can be viewed as a model-based version of reinforcement learning =-=[24,98,114]-=-; however, for the sake of consistency with the literature we will refer to it as dynamic programming (there is no learning involved at the planning level). In recent years, several researchers have p... |

360 |
Sensor fusion in certainty grids for mobile robots
- Moravec
- 1988
(Show Context)
Citation Context ...roduced two fundamental paradigms for modeling indoor robot environments: the grid-based (metric) paradigm and the topological paradigm. Grid-based approaches, such as those proposed by Moravec/Elfes =-=[19]-=- and Borenstein /Koren [2] and many others, represent environments by evenly-spaced grids. Each grid cell may, for example, indicate the presence of an obstacle in the corresponding region of the envi... |

334 | Prioritized sweeping: reinforcement learning with less data and less real time. Machine Learning 13(1):102–130
- Moore, Atkeson
- 1993
(Show Context)
Citation Context ...the value iteration update. As a result, value iteration focuses on small fraction of the grid only, hence converges much faster. Notice that the bounding box bears similarity to prioritized sweeping =-=[18]-=-. Figure 8b shows a snapshot of autonomous exploration in the environment depicted in Figure 4. The right plot, 8b, sketches the path taken during autonomous exploration. At the current point, the rob... |

329 |
A two-dimensional interpolation function for irregularly-spaced data
- Shepard
- 1968
(Show Context)
Citation Context ...lobal grid cell is projected into the local robot's coordinates, and the local occupancy value is obtained by interpolation. The interpolating function is similar in spirit to Shepard's interpolation =-=[93]-=-. It has several interesting properties, most notably it is smooth (continuous) and almost everywhere differentiable in hx robot ; y robot ; ` robot i. The key advantage of interpolating between occup... |

321 | The dynamic window approach to collision avoidance
- Fox, Burgard, et al.
- 1997
(Show Context)
Citation Context ...ually far away (=same costs). Notice that the complete exploration run shown here took less than 15 minutes. The robot moved constantly, and frequently reached a velocity of 80 to 90 cm/sec (see also =-=[3, 10]-=-). Value iteration is a very general procedure, which has several properties that make it attractive for real-time mobile robot navigation: ffl Any-time algorithm. As mentioned above, value iteration ... |

308 |
Sonar-based real-world mapping and navigation
- Elfes
- 1987
(Show Context)
Citation Context ...d approach empirically. The paper is concluded by a discussion in Section 5. 2 Grid-Based Maps The metric maps considered here are discrete, two-dimensional occupancy grids, as originally proposed in =-=[6, 19]-=- and since implemented successfully in various systems. Each grid-cell hx; yi in a map has a value attached that measures the subjective belief that this cell is occupied. More specifically, it contai... |

292 | Improving elevator performance using reinforcement learning
- Crites, Barto
- 1996
(Show Context)
Citation Context ... found in the literature on dynamic programming [4,44,83] and reinforcement learning [2,46,97]. In fact, our motion planning algorithm can be viewed as a model-based version of reinforcement learning =-=[24,98,114]-=-; however, for the sake of consistency with the literature we will refer to it as dynamic programming (there is no learning involved at the planning level). In recent years, several researchers have p... |

267 | Feudal Reinforcement Learning
- Dayan, Hinton
- 1993
(Show Context)
Citation Context ...ns using a reinforcement learning algorithm. Sub-problems are specified through "sub-goals" or certain sub-reward functions, which have to be provided manually by the human designer. Dayan a=-=nd Hinton [26]-=- proposed a hierarchical reinforcement learning architecture which recursively decomposes the state space into squares of fixed size. At each level of control, policies are generated for moving from o... |

263 | Probabilistic robot navigation in partially observable environments
- Simmons, Koenig
- 1995
(Show Context)
Citation Context ...ation (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. While grid-based maps yield more accurate control, planning with more abstract t... |

258 | Directed Sonar Sensing for Mobile Robot Navigation - Leonard, Durrant-Whyte - 1992 |

249 |
Complexity of the movers problem and generalizations
- Reif
- 1979
(Show Context)
Citation Context ... also addressed issues of efficiency. A key difficulty arises from the observation that robot motion planning, in its general definition, is worst-case exponential in the number of degrees of freedom =-=[13,84]-=-. Due to the strong focus on consistency, O()-type complexity, worst case analysis and robots with many degrees of freedom, existing cell decomposition methods usually decompose the free-space in odd ... |

236 | The parti-game algorithm for variable resolution reinforcement learning in multidimensional statespaces
- Moore, Atkeson
- 1995
(Show Context)
Citation Context ...ar to Kaelbling's approach, decomposes robot planning problems into sets of smaller problems by selecting a small number of random points can be found in [49]. Similar to Dayan and Hinton [26], Moore =-=[68]-=- recently proposed an approach for decomposing space into a set of rectangles, called parti-game. In his approach, the resolution of the decomposition is variable. It maximal along the boundary betwee... |

200 | Reinforcement learning with perceptual aliasing: The perceptual distinctions approach
- Chrisman
- 1992
(Show Context)
Citation Context ...g Finite State Automata Within the AI community, research has been conducted on general methods that can reverse-engineer (learn) finite state automata based on their input-output behavior (see e.g., =-=[3,20,78,66,72,86,87]). Finite -=-state automata (FSAs) are learned by observing the result of sequences of actions. Often, algorithms capable of learning FSAs require a pre-given "homing sequence," i.e., a sequence that res... |

192 |
An experiment in guidance and navigation of an autonomous robot vehicle
- Cox, Blanche
- 1991
(Show Context)
Citation Context ...age and drift can have devastating effects on the estimation of the robot position. Identifying and correcting for slippage and drift (odometric error) is therefore an important issue in map building =-=[7,22,85]-=-. Figures 5 and 10 give examples that illustrate the importance of position estimation in grid-based robot mapping. For example, in Figure 5a the position is determined solely based on dead-reckoning.... |

189 |
Position referencing and consistent world modeling for mobile robots
- Chatila, Laumond
- 1985
(Show Context)
Citation Context ... readings are interpreted in the context of their neighbors, which increases the accuracy of the resulting maps [99]. Occupancy grids, however, are not the only metric representation. Chatila/Laumond =-=[15]-=- proposed to represent objects by polyhedra in a global coordinate frame. Cox [23] proposed to constructs probabilistic trees to represent different, alternative models of the environment. In his work... |

189 |
Dynamic map building for an autonomous mobile robot
- Leonard, Durrant-Whyte, et al.
- 1992
(Show Context)
Citation Context ...ing and localization is significantly more difficult than either task in isolation [85]. There are several attempts to integrate localization and mapping. For example, Leonard, Durrant-Whyte, and Cox =-=[62]-=- proposed a method that interleaves localization and mapping using Kalman filters [48] for position tracking. In their experiments, however, only the mapping component of their approach is demonstrate... |

182 | Automatically generating abstractions for planning
- Knoblock
- 1994
(Show Context)
Citation Context ...ing community, such algorithms are usually referred to as cell decomposition methods [92,60]. Within Artificial Intelligence, algorithms of this type are usually referred to as abstraction algorithms =-=[43,51,89]-=-. There is a huge body of literature on cell decomposition for robot motion planning. For example, Schwartz and Sharir published a series of five seminal papers in which the motion planning problem fo... |

178 | Estimating the absolute position of a mobile robot using position probability grids - Burgard, Fox, et al. - 1996 |

171 |
an office-navigating robot
- Nourbakhsh, Powers, et al.
- 1995
(Show Context)
Citation Context ... its genuine computational simplicity, partially because landmarks appear to play a major role in human navigation [19]. Examples of successful algorithms for landmark-based localization can be found =-=[5,21,52,47,50,74,76,80,95,111]-=- and various chapters in [53]. -- Model matching. Model matching algorithms extract geometric features from the sensor readings and match those to a model of the environment in order to identify error... |

165 | Transfer of learning by composing solutions for elemental sequential tasks
- Singh
- 1992
(Show Context)
Citation Context ...approaches can roughly be divided into two classes, those that rely on a fixed decomposition, and those that decompose the state space by themselves during problem solving. -- Fixed decomposition. In =-=[63,96,110]-=- algorithms are presented that first learn solutions to sub-problems (using model-free reinforcement learning), then combine these solutions using a reinforcement learning algorithm. Sub-problems are ... |

163 |
Randomized incremental construction of Delaunay and Voronoi diagrams
- Guibas, Knuth, et al.
- 1992
(Show Context)
Citation Context ...roach, each landmark state defines a region and states other than landmark states are members of the region defined by the nearest landmark state. This approach is a version of Delaunay triangulation =-=[29,39]-=-, a family of methods that decompose the state space through Voronoi diagrams. Just like in Dayan/Hinton's and Dean/Lin's approach, Kaelbling's approach applies dynamic programming at multiple levels:... |

159 | World modeling and position estimation for a mobile robot using ultrasonic ranging - Crowley - 1989 |

158 |
Topological mapping for mobile robots using a combination of sonar and vision sensing
- Kortenkamp, Weymouth
- 1994
(Show Context)
Citation Context ...onments by evenly-spaced grids. Each grid cell may, for example, indicate the presence of an obstacle in the corresponding region of the environment. Topological approaches, such a those described in =-=[7, 14, 15, 17, 22, 35]-=-, represent robot environments by graphs. Nodes in such graphs correspond to distinct situations, places, or landmarks (such as doorways). They are connected by arcs if there exists a direct path betw... |

157 | Interaction and Intelligent Behavior
- Mataric
- 1994
(Show Context)
Citation Context ...onments by evenly-spaced grids. Each grid cell may, for example, indicate the presence of an obstacle in the corresponding region of the environment. Topological approaches, such a those described in =-=[7, 14, 15, 17, 22, 35]-=-, represent robot environments by graphs. Nodes in such graphs correspond to distinct situations, places, or landmarks (such as doorways). They are connected by arcs if there exists a direct path betw... |

155 | Randomized preprocessing of configuration space for fast path planning
- Kavraki, Latombe
(Show Context)
Citation Context ... with excessive degrees of freedom that, similar to Kaelbling's approach, decomposes robot planning problems into sets of smaller problems by selecting a small number of random points can be found in =-=[49]-=-. Similar to Dayan and Hinton [26], Moore [68] recently proposed an approach for decomposing space into a set of rectangles, called parti-game. In his approach, the resolution of the decomposition is ... |

147 |
Occupancy Grids: A probabilistic framework for robot perception and navigation
- Elfes
- 1989
(Show Context)
Citation Context ...oaches to map learning with mobile robots were either metric or topological (sometimes enriched by metric information). While the idea of integrating metric and topological representations is not new =-=[15,32,31]-=-, it has not yet been demonstrated that this can actually be done robustly in environments that are significantly larger than the perceptual range of a robot's sensors. Thus, the major contribution of... |

139 |
A distributed model for mobile robot environment-learning and navigation
- Matarić
- 1990
(Show Context)
Citation Context ...Each grid cell may, for example, indicate the presence of an obstacle in the corresponding region of the environment. Topological approaches, such a those proposed by Kuipers/Byun, Mataric and others =-=[34,55,58,65,81,105,112,115]-=-, represent robot environments by graphs. Nodes in such graphs correspond to distinct situations, places, or landmarks (such as doorways). They are connected by arcs if there exists a direct path betw... |

133 | Efficient training of artificial neural networks for autonomous navigation
- Pomerleau
- 1991
(Show Context)
Citation Context ... in the building in which the software was originally developed. Even though time was short, the neural network could quickly be retrained to accommodate this new situation. Others, such as Pomerleau =-=[23]-=-, also report a significant decrease in development time of integrated robotic systems through the use of machine learning algorithms. 2. Multiple sensor reading are interpreted simultaneously. Most c... |

127 | Mobile robot localization using landmarks
- Betke, Gurvits
- 1997
(Show Context)
Citation Context ... its genuine computational simplicity, partially because landmarks appear to play a major role in human navigation [19]. Examples of successful algorithms for landmark-based localization can be found =-=[5,21,52,47,50,74,76,80,95,111]-=- and various chapters in [53]. -- Model matching. Model matching algorithms extract geometric features from the sensor readings and match those to a model of the environment in order to identify error... |

113 | The Mobile Robot RHINO
- Buhmann, Burgard, et al.
- 1995
(Show Context)
Citation Context .... Thrun and A. Bucken Both approaches to robot mapping exhibit orthogonal strengths and weaknesses. Occupancy grids are considerably easy to construct and to maintain even in large-scale environments =-=[3]-=-. Since the intrinsic geometry of a grid corresponds directly to the geometry of the environment, the robot's position within its model can be determined by its position and orientation in the real wo... |

105 | Integrating grid-based and topological maps for mobile robot navigation - Thrun, Bücken - 1996 |