Since its beginning, the robotics field had tried to increase the interaction between robots and their environment. The first robots were only able to perform repetitive tasks. These robots were hardwired so there was no possibility of changing their behaviour. In order to introduce interaction between humans (such as the programmer) and the robot, people introduced robots that were able to be reprogrammed.
The problem with these robots is that they must be programmed for each new situation. Indeed, a programmer cannot design all the possible situations. This is especially true for robots which have to operate in unconstrained environments where people move about too.
To be efficient, a robot must interact with its environment. It has to be able to react correctly to unknown events that can occur. Learning by experience is then a useful feature we can implant on a robot. In this thesis we are particularly interested in corridor following and egomotion estimation using visual flow. These are two features that are needed to navigate correctly using a map provided by the user.
Egomotion estimation is also needed as feedback for odometry. Indeed, wheel encoders are often not accurate. Imagine the difference between a robot driving a long distance on a road and the same robot driving on the sand. In the first case, the robot is moving forward and the odometry system is approximately recording the right distance. In the second case, the robot just slips, but the odometry system records a long distance. This recorded distance is wrong because it should be approximately zero. The slippage confuses the odometry system because it involves an unknown friction coefficient. This situation is quite frequent in indoors environments too. If a robot is stuck against a wall, its wheels turn but the robot does not move.
On some robots, we cannot use wheel encoders to give us an idea of the distance travelled so far. Basic examples are helicopters, submarines, legged robots and any robots which do not have wheels. In these situations, egomotion estimation can be essential.