We are working on robot perception challenges, especially on how to use vision to understand 3D scenes by solving problems, such as object detection and 6D object pose estimation. In the past we have also dealt with other perception challenges, such as bearing-only navigation and localization, as well as localization based on wireless signal strength.
A requirement in order to be able to plan the motion of a robotic arm in a cluttered environment is to be able to detect the objects in the robot’s vicinity and their 6D pose (i.e., location and orientation). The goal of our work is to build the capability to identify accurate pose estimates for objects in cluttered scenarios. Particularly, we have been working on 1) developing intelligent techniques to autonomously generate labeled datasets for training object recognition pipelines, and 2) developing search-based algorithms for scene estimation, given RGBD data and 3D CAD models of objects.
Oftentimes, robot navigation schemes rely on having accurate distance information in the form of laser-range scanners or sonar. This work focuses on navigation using only bearing information, rather than using distance information. The robot can accurately determine the relative bearing of landmarks in its environment using a panoramic camera. Using this bearing information, the robot is able to execute a long and complex trajectory in order to complete some desired task and then return to its original position with a high degree of accuracy. This work focuses on the theoretical guarantees provided under an ideal model and proves navigability in two-dimensional workspaces under this model.
This work focuses on studying the problem of bearing-only Simultaneous Localization and Mapping (SLAM) for robotic systems using only bearing information. A deep and wide study into different approaches to the problem is given, investigating methods such as the Extended Kalman Filter (EKF), Expectation Maximization (EM), and Particle Filtering. This work shows that particle filters work particularly well, especially when extra steps are taken to improve their robustness to outliers.