A robot receives raw sensor data from its sensors. It has to map those measurements into an internal representation to formalize this data. This process is called robotic perception. This is a difficult process in general the sensors are noisy and the environment is partially observable, unpredictable, and often dynamic.
Good representation should meet three criteria: they should
• Contain enough information for the robot to make a right decision.
• Be structured in a way that it can be updated efficiently.
• Be natural, meaning that internal variables correspond to natural state variables in the physical world.
Filtering and updating the belief state is not covered here as it was covered in earlier presentations. Some topics are Kalaman filters and dynamic Bayes nets.
A very generic perception task is localization. It is the problem of determining where things are. Localization is one of the most pervasive perception problems in robotics. Knowing the location of objects in the environment that the robot has to deal with is the base for making any successful interaction with the physical world. There are three increasingly difficult flavors of localization problems:
• Tracking – if the initial state of the object to be localized is known you can just track this object.
• Global localization – the initial location of the object is unknown you first have to find the object.
• Kidnapping – this is the most difficult task.
No comments:
Post a Comment