Action Filename Description Size Access License Resource Version
Show more files...


The present work tackles integration in mobile robotics. Integration is often considered to be a mere technique, unworthy of scientific investigation. On the contrary, we show that integrating capabilities in a mobile robot entails new questions that the parts alone do not feature. These questions reflect the structure of the application and the physics of the world. We also show that a successful integration process transforms the parts themselves and allows to scale up mobile-robot intelligence in real-world applications. In Chapter 2 we present the hardware. In Chapter 3, we show that building a low-level control architecture considering the mechanic and electronic reality of the robot improves the performances and allows to integrate a large number of sensors and actuators. In Chapter 4, we show that globally optimising mechatronic parameters considering the robot as a whole allows to implement slam using an inexpensive sensor with a low processor load. In Chapter 5, we show that based on the output from the slam algorithm, we can combine infrared proximity sensors and vision to detect objects and to build a semantic map of the environment. We show how to find free paths for the robot and how to create a dual geometric-symbolic representation of the world. In Chapter 6, we show that the nature of scenarios influences the implementation of a task-planning algorithm and changes its execution properties. All these chapters contribute results that together prove that integration is a science. In Chapter 7, we show that combining these results improves the state of the art in a difficult application : autonomous construction in unknown environments with scarce resources. This application is interesting because it is challenging at multiple levels : For low-level control, manipulating objects in the real world to build structures is difficult. At the level of perceptions, the fusion of multiple heterogeneous inexpensive sensors is not trivial, because these sensors are noisy and the noise is non-Gaussian. At the level of cognition, reasoning about elements from an unknown world in real time on a miniature robot is demanding. Building this application upon our other results proves that integration allows to scale up machine intelligence, because this application shows intelligence that is beyond the state of the art, still only combining basic components that are individually slightly behind the state of the art.