Multiagent Intelligent Surveillance System (MAIS-S)
This project, started in 2010 as a partnership
between research teams at ISR/IST, INESC-ID, Carnegie Mellon
University and Observit, studies the problem of decentralized
multiagent cooperation, applied to the setting of surveillance
networks, operated by mixed human/robot teams. This involves research in planning and autonomous decision-making , communication networks, and computer
vision. Visit the project's website here.
My role in the project has been to model real-world systems of networked cameras and robots as multiagent Partially Observable Markov Decision Processes
(POMDPs), and use these models to obtain approximately optimal action plans for the multiagent team. To this end, my contributions have so far been the following:
·I've developed techniques to minimize communication usage in multiagent POMDPs. Communication in real-world multi-robot systems is often essential to reduce
planning complexity, but in many scenarios, communication is not free;
·I've extended the POMDP framework to handle systems with multiple asynchronous agents. The resulting theory allows real-time event-driven systems, such as typical surveillance/monitoring systems, to be efficiently modeled as POMDPs, taking into account false positive/false negative event detections;
·I'm currently developing ROS packages that will ease the deployment of MDP/POMDP-based solutions to teams of real robots. These packages will be made public after the conclusion of my PhD work. Its main features include: interaction with graphical modeling software for MDPs/POMDPs (OpenMarkov); interaction with the MADP toolbox, for efficient planning over POMDP models; an easy-to-use API for the implementation of synchronous/asynchronous (PO)MDP controllers, providing abstractions for states/actions/observation in general multiagent systems.