Tham khảo tài liệu 'mobile robots perception & navigation part 8', kỹ thuật - công nghệ, cơ khí - chế tạo máy phục vụ nhu cầu học tập, nghiên cứu và làm việc hiệu quả | Global Navigation of Assistant Robots using Partially Observable Markov Decision Processes 271 . Action and observation uncertainties Besides the topology of the environment it s necessary to define some action and observation uncertainties to generate the final POMDP model transition and observation matrixes . A first way of defining these uncertainties is by introducing some experimental hand-made rules this method is used in Koenig Simmons 1998 and Zanichelli 1999 . For example if a Follow action aF is commanded the expected probability of making a state transition F is 70 while there is a 10 probability of remaining in the same state N no action a 10 probability of making two successive state transitions FF and a 10 probability of making three state transitions FFF . Experience with this method has shown it to produce reliable navigation. However a limitation of this method is that some uncertainties or parameters of the transition and observation models are not intuitive for being estimated by the user. Besides results are better when probabilities are learned to more closely reflect the actual environment of the robot. So our proposed learning module adjusts observation and transition probabilities with real data during an initial exploration stage and maintains these parameters updated when the robot is performing another guiding or service tasks. This module that also makes easier the installation of the system in new environments is described in detail in section 8. 5. Navigation System Architecture The problem of acting in partially observable environments can be decomposed into two components a state estimator which takes as input the last belief state the most recent action and the most recent observation and returns an updated belief state and a policy which maps belief states into actions. In robotics context the first component is robot localization and the last one is task planning. Figure 5 shows the global navigation architecture of the SIRAPEM
