Grasping unknown semi-transparent objects is a particularly challenging tasks, as typical 3D sensors, including LiDAR, structured light, and vision fail modelling semi-transparent objects. Our approach leverages a state-of-the-art method and a novel dataset to model 3D scene geometry and provides valid grasps, even in the presence of semi-transparent objects. The method is lightweight, and even though it uses vision sensors primarily, it can also leverage others such as RGBD or LiDAR.
April 2026

Learning to Grasp (L2G) is an efficient end-to-end learning strategy to generate 6-DOF parallel-jaw grasps starting from a partial point cloud of an object. Our approach does not exploit any geometric assumption, it is instead guided by a principled multi-task optimization objective that generates a diverse set of grasps by combining contact point sampling, grasp regression, and grasp evaluation. This video shows real-world experiments with grasps predicted from our L2G model and the baseline GPNet.
March 2022

The video presents Robust Depth, a deep learning-based model to generate depth images from monocular images.
October 2023

The video shows the output of our Graph Neural Network estimating the level of social compliance while the robot is moving in a simulated scenario.
September 2019

As the video on the right-hand side, this video shows our robot Shelly in a fetch-and-deliver task. The two arms previously found in Shelly have been replaced by a single arm with zero backslash and more precision. It must be noted that in this case the video is played in real time: without backslash in the joints visual servoing is unnecessary.
October 2016

This video shows how our robot Shelly fetches an object from a table and delivers it to a human. It was recorded for the oral defense of the Ph.D. thesis of our former student and colleague Luis V. Calderita. One of the biggest difficulties (in addition to the fact that, thanks to AGM, the robot is capable of maintaining a model of the environment, the objects and humans in it) was to achieve this task without a fine kinematic calibration, using visual feedback instead. Although calibrating the robot would make it perform the task faster, avoiding the need of a proper calibration allowed us to demonstrate how robust can a robot be using visual servoing.
February 2016

This video shows our robot Gualzru finding a coffee mug for us. As an intermediate step it also models the room in which it is located and finds a table. This video was recorded as part of my Ph.D. thesis.
The robot uses AGM and RoboComp.
June 2013