Michele Ciciolla

Gazebo simulation

Building a simulated robot for testing algorithms from home

The first task i've been ask to take care was to build a reliable simulation to test our algorithm on. Since we did not have a real robot to work on, this task was fundamental for the whole team. I've started following tutorials on Gazebo and ROS best pratictes online and i've manage to deliver a reliable simulation of our future robot.

Navigation stack setup and tuning

Scouting the best opensource planner and localizer

To deliver a prototype as fast as possible to the market we could not write a slam algorithm from scratch much less a planning controller. For that reason my task has been to scout SLAM suites on the opensource network in ROS. I've found several interesting solutions for us, but at the end RTABMAP demonstrated to be the most accurate. Using this algorithm we manage to build accurate 3D maps of the environment aswell as providing a good localization in medium-wide areas.
Planner choice has been quite straight forward since TEB is considered one of the best planner around in ROS.

Computer vision pipeline for trash picking

Get the position of a trash in robot coordinates and pick it

I built from scratch a computer vision pipeline being able to compute the position of an object in robot coordinates and plan a trajectory to pick it with the brushes. The input of this pipeline is the bounding box information provided by the our custom YOLO network and those bounding box information are used to compute the pixel coordinates of the object in the frontal camera image. Then it comes a pointcloud filtering and segmentation extracting the 3D location of the trash according to the robot center. The final step is to command the planner to send the robot to that location.

Detecting downstairs using pointcloud

Avoiding our robot to fall downstairs

For safety reason we had to build some controller checking that the planned trajectory is safe and for such reason i've been ask to think about a solution. Since in the previous project i've worked a bit with pointclouds, my solution came quite easy: this controller tryes to segment a ground plane of points from the pointcloud and send alarm once it cannot.

Sensors and node fault detection

Reporting sensors fault reading from topics

ROS1 is not ideal and sometimes it fails to update and send topics data along the wire. That's why we had to understand when this happens: this topic and node manager checks in a loop every topic you want to track and report an alarm for each of them not responding for a certain amount of time.

Master Thesis Project

An approach to camera calibration in Industrial Robotics using HexSight

Quality inspection tasks require the camera to be calibrated. In the industrial robotics context the process of camera calibrations is usually performed manually by a human operator, so the goal of this thesis was to develop a automatic way to perform the calibration using an ABB robot, a camera and a machine vision software (HexSight). The results obtained with this process demonstrated that automatic camera calibration can achieve sub-milimetric accuracy aswell.

Mobile Robot Navigation and Mapping (SLAM) using Range Sensor

Here we simulate a mobile robot simultaneously mapping landmark positions and its trajectory while sensing them. The robot has a range sensor which can obtain the distance between itself and a landmark. In this project correspondences are taken as granted, so i focused more on the optimization algorithm.

Visual/Force Control of a surgical robot (DaVinci)

The aim of this project was to implement a hybrid visual and force controller for the DaVinci robot in a simulated environment using MATLAB and VREP-CoppeliaSim.

Consensus strategy platooning of autonomous vehicles

Autonomous driving is the future of private and commercial tranport and the paper implemented in this project shows a potential distributed control of several vehicles one behind the other. The first vehicle is the leader of the platoon and it chooses the velocity the group has to maintain. The vehicles behind him follow it keeping a custom safety distance each other. Tests performed successfully even in presence of communication delays and suddenly brakes of the leader.

CNN Covid Detector

A convolutional neural network was trained here to classify positive and negative COVID patients based on their X-Ray image analysis. For this project me and my collegues we have chosen to use transfer learning to obtain a better performance in classification. We've trained the last layers of a ResNet101 on a small labeled dataset we've found in the web managing to obtain a 70% classification accuracy.