## Unsupervised Learning of Probabilistic Models for Robot Navigation (1996)

Venue: | in Proceedings of the IEEE International Conference on Robotics and Automation |

Citations: | 86 - 4 self |

### BibTeX

@INPROCEEDINGS{Koenig96unsupervisedlearning,

author = {Sven Koenig and Reid G. Simmons},

title = {Unsupervised Learning of Probabilistic Models for Robot Navigation},

booktitle = {in Proceedings of the IEEE International Conference on Robotics and Automation},

year = {1996},

pages = {2301--2308}

}

### Years of Citing Articles

### OpenURL

### Abstract

Navigation methods for office delivery robots need to take various sources of uncertainty into account in order to get robust performance. In previous work, we developed a reliable navigation technique that uses partially observable Markov models to represent metric, actuator, and sensor uncertainties. This paper describes an algorithm that adjusts the probabilities of the initialMarkov model by passively observing the robot's interactions with its environment. The learned probabilities more accurately reflect the actual uncertainties in the environment, which ultimately leads to improved navigation performance. The algorithm, an extension of the Baum-Welch algorithm, learns without a teacher and addresses the issues of limited memory and the cost of collecting training data. Empirical results show that the algorithm learns good Markov models with a small amount of training data. 1 Introduction Navigation methods for office delivery robots need to take various sources of uncertainty i...

### Citations

922 | An introduction to Hidden Markov Models
- Rabiner
- 1986
(Show Context)
Citation Context ... and must deal with the fact that collecting data is time consuming. Our POMDP learning algorithm addresses all these concerns. It is an unsupervised, passive method based on the Baum-Welch algorithm =-=[11]-=-, a simple expectationmaximization algorithm for learning POMDPs from observations. To enable the algorithm to run on-board the robot, we have extended the Baum-Welch algorithm to use a floating windo... |

268 | Probabilistic robot navigation in partially observable environments
- Simmons, Koenig
- 1995
(Show Context)
Citation Context ...avier mobile robot (Figure 1) that explicitly represents uncertain metric information (e.g. "corridor X is between 2 and 9 meters long"), actuator (dead-reckoning) uncertainty, and sensor un=-=certainty [13]-=-. The technique uses partially observable Markov decision process (POMDP) models to estimate the position of the robot in the form of probability distributions. Experience with the technique has shown... |

173 |
Dervish: An office-navigating robot
- Nourbakhsh, Powers, et al.
- 1995
(Show Context)
Citation Context ...about the lengths of corridors and improving the accuracy of the actuator and sensor models. Our approach is fairly general, and is applicable to other robot navigation systems that use POMDPs (e.g., =-=[9]-=-). We desire that the learning be unsupervised and passive. Unsupervised learning means that the robot gets no information from a teacher, such as where it really is or what it really observed. This i... |

82 | A robust, qualitative method for robot spatial learning
- Kuipers, Byun
- 1988
(Show Context)
Citation Context ...reach a given goal location. Approaches whose properties have been analyzed formally, for example, include [5], [10], [12], [7], and [15]. Approaches that have been demonstrated experimentallyinclude =-=[6]-=-, [8], and [14]. The approaches of [1] and [2] learn Markov models of the environment, as we do, but they use active exploration, while our approach is passive. The learning approach of [4] does use a... |

61 |
Error correction in mobile robot map learning
- Engelson, McDermott
- 1992
(Show Context)
Citation Context ...llyinclude [6], [8], and [14]. The approaches of [1] and [2] learn Markov models of the environment, as we do, but they use active exploration, while our approach is passive. The learning approach of =-=[4]-=- does use a passive learning approach, but it learns a topological map only. These approaches also differ from our learning method in that they learn their models from scratch. 8 Conclusion This paper... |

58 |
Dynamic path planning in sensor-based terrain acquisition
- Lumelsky, Mukhopadhyay, et al.
- 1990
(Show Context)
Citation Context ...rs or sensors) with the goal either to map the environment completely or to reach a given goal location. Approaches whose properties have been analyzed formally, for example, include [5], [10], [12], =-=[7]-=-, and [15]. Approaches that have been demonstrated experimentallyinclude [6], [8], and [14]. The approaches of [1] and [2] learn Markov models of the environment, as we do, but they use active explora... |

58 | Exploration and model building in mobile robot domains
- Thrun
- 1993
(Show Context)
Citation Context ...goal location. Approaches whose properties have been analyzed formally, for example, include [5], [10], [12], [7], and [15]. Approaches that have been demonstrated experimentallyinclude [6], [8], and =-=[14]-=-. The approaches of [1] and [2] learn Markov models of the environment, as we do, but they use active exploration, while our approach is passive. The learning approach of [4] does use a passive learni... |

56 |
A mobile robot exploration algorithm
- Zelinsky
- 1992
(Show Context)
Citation Context ...sors) with the goal either to map the environment completely or to reach a given goal location. Approaches whose properties have been analyzed formally, for example, include [5], [10], [12], [7], and =-=[15]-=-. Approaches that have been demonstrated experimentallyinclude [6], [8], and [14]. The approaches of [1] and [2] learn Markov models of the environment, as we do, but they use active exploration, whil... |

44 |
Baum’s forward-backward algorithm revisited
- Devijver
- 1985
(Show Context)
Citation Context ...orithm uses the following dynamic programming approach ("forward-backward algorithm"), that applies Bayes' rule repeatedly, to calculate the scaling factors, alpha values, and beta values ef=-=ficiently [3]: A1. Set -=-scale 1 := P s2S [p(o 1 js)P(s 1 = s)]. A2. Set ff 1 (s) := p(o 1 js)P(s 1 = s)=scale 1 for all s 2 S. A3. For t := 1 to T \Gamma 1 ("forward propagation") (a) Let temp t (s) = P s 0 2S [p(s... |

41 | Inferring finite automata with stochastic output functions and an application to map learning
- Dean, Angluin, et al.
- 1992
(Show Context)
Citation Context ...properties have been analyzed formally, for example, include [5], [10], [12], [7], and [15]. Approaches that have been demonstrated experimentallyinclude [6], [8], and [14]. The approaches of [1] and =-=[2]-=- learn Markov models of the environment, as we do, but they use active exploration, while our approach is passive. The learning approach of [4] does use a passive learning approach, but it learns a to... |

32 | Coping with uncertainty in map learning
- Basye, Dean, et al.
- 1989
(Show Context)
Citation Context ...s whose properties have been analyzed formally, for example, include [5], [10], [12], [7], and [15]. Approaches that have been demonstrated experimentallyinclude [6], [8], and [14]. The approaches of =-=[1]-=- and [2] learn Markov models of the environment, as we do, but they use active exploration, while our approach is passive. The learning approach of [4] does use a passive learning approach, but it lea... |

28 |
Environment learning using a distributed representation
- Mataric
- 1990
(Show Context)
Citation Context ... a given goal location. Approaches whose properties have been analyzed formally, for example, include [5], [10], [12], [7], and [15]. Approaches that have been demonstrated experimentallyinclude [6], =-=[8]-=-, and [14]. The approaches of [1] and [2] learn Markov models of the environment, as we do, but they use active exploration, while our approach is passive. The learning approach of [4] does use a pass... |

28 |
A “retraction” method for learned navigation in unknown terrains for a circular robot
- Rao, Stoltzfus, et al.
- 1991
(Show Context)
Citation Context ...ctuators or sensors) with the goal either to map the environment completely or to reach a given goal location. Approaches whose properties have been analyzed formally, for example, include [5], [10], =-=[12]-=-, [7], and [15]. Approaches that have been demonstrated experimentallyinclude [6], [8], and [14]. The approaches of [1] and [2] learn Markov models of the environment, as we do, but they use active ex... |

19 |
Robot navigation in unknown terrains using learned visibility graphs. Part I: The disjoint convex obstacle case
- Oommen, Iyengar, et al.
- 1987
(Show Context)
Citation Context ...fect actuators or sensors) with the goal either to map the environment completely or to reach a given goal location. Approaches whose properties have been analyzed formally, for example, include [5], =-=[10]-=-, [12], [7], and [15]. Approaches that have been demonstrated experimentallyinclude [6], [8], and [14]. The approaches of [1] and [2] learn Markov models of the environment, as we do, but they use act... |

11 | Robot navigation algorithms using learned spatial graphs
- Iyengar, Jorgensen, et al.
- 1986
(Show Context)
Citation Context ...g perfect actuators or sensors) with the goal either to map the environment completely or to reach a given goal location. Approaches whose properties have been analyzed formally, for example, include =-=[5]-=-, [10], [12], [7], and [15]. Approaches that have been demonstrated experimentallyinclude [6], [8], and [14]. The approaches of [1] and [2] learn Markov models of the environment, as we do, but they u... |