Title: Multi-Motion Segmentation for Manipulating Unknown Objects Abstract: Robotic grasping tasks often depend on prior visual and geometric models of manipulated objects, to identify their location in 3D and to select proper grasping poses respectively. While high-resolution 3D scanning and deep learning segmentation methods make it easier than ever to create large sets of known object models, collecting these models is tedious and scenario-specific, as it becomes infeasible to cover the huge variety of objects that a robot might encounter in new scenarios.
Instead of modelling all possible objects a-priori, we motivate to rely on motion cues to segment and model objects online and thus let the robot identify the objects of interest in the current scenario. To obtain a dense geometric object representation that is suitable for grasp planning and robust to large camera and object movements, we propose to combine sparse but robust feature points with the dense segmentation from optical flow and depth data.
Title: Multi-contact teleoperation for humanoid, bimanual arms and application to the FAIR-SPACE project
Abstract: This work has been developed for the FAIR-SPACE project aiming to apply robotic technologies to space. I will present a method to safely teleoperate legged robots in a multi-contact environment for loco-manipulation tasks and under the quasi-static motion assumption. This method is extended to bimanual arms and I will describe the architecture and implementation of the teleoperation use case developed for the FAIR-SPACE project.