Informal Meeting with Alexandre Albore on Robotic Localization Issues

In an informal meeting, Alexandre Albore, PhD student in the Artificial Intelligence Group at the UPF, introduced me to the field of robotic localization. Alex is focused on the theoretical aspects of planning in AI, that is creating sequences of actions, possibility conditioned by observations, that bring a system from an initial state to a goal. We have in common our experience that imperfect observations within uncertain domains and dynamic world often challenges planning (for robots and humans). The difference between CatchBob! and his robots is that I do not give players information about the location data imprecision, while such data are required by robots.

He explained me his use of Monte Carlo localization (i.e. Markov localization, particle distribution) and Kalman filter in robotics.

We share on the multiple issues inherent to robotic localization. Robots most often use lasers (50m coverage, problems with windows) and sonars (10m coverage, echo and filter problems). New localization system now use limunosity (I am reading something on this for context-aware wearable computing). Often an integration of different localization techniques is uses (fusion). In the context of Monte Carlo localization, observations are used to disambiguate previous inferences. Issues are that there might be too much or not enough information.

He advised me to have a look at the work of: Sebastien Thurn Sven Koenig, Vadim Bulitko, Illah Nourbakhsh, Human-robot interaction, some work on localization and Eric Beaudry.

Relation to my thesis: Real-life AI applications are often characterized by uncertainty, dynamic changes of the world, and limited knowledge available a priori. As a result, researchers from several AI areas have recently invested much effort into methods suitable for domains with such kinds of incomplete information.