Multi-camera intelligent space for a robust, fast and easy deployment of proactive robots in complex and dynamic environments

Abstract

One of the current challenges in robotics is the integration of robots in everyday environments. However, it is difficult to achieve this with stand-alone robots that use only the information provided by their own sensors (on-board sensors). In this thesis, we will explore the use of intelligent spaces (i.e. spaces where many sensors and intelligent devices are distributed and which provide information to the robot), to get robots operating in complex environments in a short period of time. Our proposal is to build an intelligent space that allows an easy, fast, and robust deployment of robots in different environments. This solution must allow robots to move and operate efficiently in unknown environments, and it must be scalable to the number of robots and other elements. Our intelligent space will consist of a distributed network of intelligent cameras and autonomous robots. The cameras will detect situations that might require the presence of the robots, inform them about these situations, and also support their movement in the environment. The robots, on the other hand, will navigate safely within this space towards the areas where these situations happen. With this proposal, our robots are not only able to react to events that occur in their surroundings, but to events that occur anywhere. As a consequence, the robots can react to the needs of the users regardless of where the users are. This will look as if our robots are more intelligent, useful, and have more initiative. In addition, the network of cameras will support the robots on their tasks, and enrich their environment models. This will result on a faster, easier and more robust robot deployment and operation. In this thesis, we will explore two alternatives, regarding how the intelligence is distributed among the agents: collective intelligence and centralised intelligence. Under the collective intelligence paradigm, intelligence is fairly distributed among robots and cameras. Global intelligence arises from the interaction among individual agents, and there is not a central agent that handles most decision making. This is somehow similar to self-organization processes that are usually observed in nature, where there is no hierarchy nor centralisation. In this case, we assume that it is possible to get robots operating in a priori unknown environments when their behaviour emerges from the interaction amongst an ensemble of independent agents (cameras), that any user can place in different locations of the environment. These agents, initially identical, will be able to observe human and robot behaviour, learn in parallel, adapt and specialize in the control of the robots. To this extent, our cameras will be able to detect and track robots and humans robustly, to discover their camera neighbours, and to guide the robot navigation through routes of these cameras. Meanwhile, the robots must only follow the instructions of the cameras and negotiate obstacles in order to avoid collisions. On the other hand, under the centralised intelligence paradigm, one type of agent will be assigned much more intelligence than the rest. Therefore, this agent will make most decision making and coordination, and its performance will have a higher importance than that of other agents. To explore this paradigm, in this thesis, the role of central agent will be played by the robot agent, and most of this intelligence will be devoted to the task of self-localisation and navigation. In this regard, we have performed an experimental study about the strengths and weaknesses of different information sources to be used for the task of robot localisation. The study has shown that no source performs well in every situation, but the combination of complementary sensors may lead to more robust localisation algorithms. Therefore, we have developed a robot localisation algorithm that combines the information from multiple sensors. This algorithm is able to provide robust and precise localisation estimates even in situations where singlesensor localization techniques usually fail. It can fuse the information of an arbitrary number of sensors, even if they are not synchronised, work at different data rates, or if some of them stop working. We have tested our algorithm with the following sensors: a 2D laser range finder, a magnetic compass, a WiFi reception card, a radio reception card (433 MHz band), the network of external cameras, and a camera mounted in the robot. We have also designed wireless transmitters (motes) and we have studied the performance of our positioning algorithm when they are able to vary their transmission power. Through an experimental study, we have demonstrated that this ability tends to improve the performance of a wireless positioning system. This opens the door for future improvements in the line of active localisation. Under this paradigm, the robot would be able to modify the transmission power of the transmitters in order to discard localisation hypotheses proactively. Our proposal is a generic solution that can be applied to many different service robot applications. As an specific example of application, we have integrated our intelligent space with a general purpose guide robot that we have developed in the past. This robot is aimed to operate in different social environments, such as museums, conferences, or robotics demonstrations in research centres. Our robot is able to detect and track people around him, follow an instructor around the environment, learn routes of interest from the instructor, and reproduce them for the visitors of the event. Moreover, the robot is able to interact with humans using gesture recognition techniques and an augmented reality interface.