Autonomous exploration with a vision-based mobile robot

By | March 2, 2007

In recent years, the vast majority of mobile robots have been constructed to rely heavily on a laser sensor for perceiving their surrounding environment. Laser is very accurate when it comes to measuring the distance to objects making it possible for mobile robots to navigate and map large and highly dynamic environments. On the other hand, laser sensors can only scan along a single plane and often fail to perceive large objects such as tables; a laser sensor would only detect the legs of the table.

Stereo vision is a much better sensor for use with mobile robots because not only can it measure distance to objects but image data can also be used for other tasks such as localization, object and people recognition. Researchers have avoided using stereo vision on their mobile robots because in the past computers were not fast enough to process all the data and also because we did not have the basic algorithms to make computer vision work.

Recent advances in computer vision have made it possible to construct visually guided mobile robots that are as capable as the robots depending on laser. For example, Rob Sim and his colleagues at the University of British Columbia, Canada, have developed a mobile robot that can estimate its location and construct a map of its environment using only visual feedback.

Last summer, Rob demonstrated the capabilities of his robot when he published a paper showing it autonomously exploring a large office-like environment correctly estimating its position using a Rao-Blackwellised particle filter (RBPF) and in addition constructing an occupancy grid map. An occupancy grid is a 2D plan of the environment such that the area occupied by stationary objects is marked along with the space that is free and can be traversed by the robot.

Here, I have included a video that shows the map construction process. The robot travels more than 100 meters mapping an area more larger than 400 square meters. The robot is shown with a yellow rectangle while the path it follows is in yellow. The red path displays the direction that the robot decides to move in order to maximize the accuracy of the learned map and in blue is the raw odometry or better the robot’s uncorrected trajectory. Black pixels corresponding to occupied space, i.e., walls and furniture, while white pixels denote empty space. Grey pixels denote space that the robot has not observed because for example it is outside the camera’s field of view or it is space hidden behind another object. There are more movies showing the mapping algorithms in action here.