Carlos Costa

Robot self-localization in dynamic environments

​

   

  Mobile robot platforms capable of operating safely and accurately in dynamic environments can have a multitude of applications, ranging from simple delivery tasks to advanced assembly operations. These abilities rely heavily on a robust navigation stack, which requires a stable and accurate localization system.

     This dissertation proposes an efficient, modular, extensible and easy to configure 3/6 DoF localization system, capable of operating on a wide range of mobile robot platforms and environments. It is able to reliably estimate the global position using feature matching and is capable of achieving high accuracy pose tracking with a range of point cloud registration algorithms. It can use several point cloud sensing devices (such as LIDARs or RGB-D cameras) and by using reference maps it does not require ambient modification. Moreover, it can update the localization map at runtime and dynamically adjust its operation rate based on the predicted robot velocity in order to use the minimum amount of hardware resources. It also offers a detailed analysis of each pose estimation, providing information about the percentage of registered inliers, the root mean square error of the inliers, the angular distribution of the inliers and outliers, the pose corrections that were performed in relation to the expected position and in case of initial pose estimation it also gives the distribution of the acceptable initial poses, which can be very valuable information for a navigation supervisor when the robot is in ambiguous areas that are very similar in different parts of the known environment.

   The ROS implementation was tested in several dynamic indoor environments using three mobile robot platforms equipped with LIDARs and RGB-D cameras. Overall tests using sensor data from simulation and retrieved from the robot platforms performed in a high end laptop with an Intel Core i7 3630QM processor, 16GB DDR3 of memory and NVIDIA GTX680M graphics card, demonstrated high accuracy in complex dynamic environments, with less than 1-2 cm in translation error and less than 1-3 degrees in rotation error (in 3 and 6 DoF respectively). Execution times ranged from 5 to 35 milliseconds in both 3 and 6 DoF.

     The sub centimeter accuracy achieved by the proposed localization system along with the dynamic map update capability and the need of no artificial landmarks will allow the fast deployment of mobile robot platforms capable of operating safely and accurately in cluttered environments. Moreover, the robustness to dynamic objects will grant the possibility to use robots as coworkers, helping humans perform their work more efficiently and thus reducing the overall production costs.