Work

Highlights

• Designed the camera assembly, configuration and calibration workflow used during WeMo production and commissioning.
• Built automated intrinsic and extrinsic calibration pipelines plus tooling to manage per-robot calibration artefacts.
• Defined validation procedures for relative positioning, targeting better than 5 mm positional error and 0.5 degree angular error.
• Reduced manual calibration effort and improved repeatability across fleet deployments.

Highlights

• Designed the Relative Move To capability for short-range autonomous docking and alignment using onboard cameras and fiducial targets.
• Used a teaching-by-showing approach so desired poses could be learned once and reused across the fleet instead of reteaching every target on every robot.
• Defined the goal model and factory integration flow for station- and payload-specific positioning targets, including cluster-mode pallet delivery.
• Calibration cut cross-robot pose-transfer error by roughly 8x in translation and more than 15x in orientation, despite lens and sensor variability, camera mounting differences and manufacturing tolerances.
• Validated the method with motion-capture-based accuracy campaigns, including multi-robot cluster tests where poses learned on one robot still achieved about 1.8 to 2.8 mm translation accuracy and about 0.04 to 0.1 degrees angular accuracy.
• Enabled reliable coupling to charging stations and factory interfaces after coarse navigation.

Highlights

• Trained and deployed convolutional neural networks for object detection and semantic segmentation on the Roborace testbed.
• Contributed to the perception and obstacle-avoidance stack validated on track with the DevBot 2.0 autonomous race car.
• Worked alongside Data:Lab Munich data scientists on machine-learning algorithms optimised for high-speed, dynamic racing conditions.

Highlights

• Led the ML and computer-vision work for a real-time passenger-information system combining onboard and station cameras.
• Used CNN classification for train presence, coach occupancy estimation and alighting-flow estimation.
• Used detection and tracking for door open/close events, people detection and people counting.
• Built a random-forest model to predict how many passengers would exit each train.
• Deployed successfully at the Re Umberto station in Turin from May 2019.

Highlights

• Perception system for obstacle and free-navigable-space detection.
• Human-machine shared control with driver assistance for obstacle avoidance.

Highlights

• Defined the hardware and software architecture for the autonomous ground module, including cameras, lidars, radars, ultrasonics, GNSS/IMU, NVIDIA PX2 and MicroAutoBox.
• Perception work spanning 2D and 3D obstacle detection, semantic segmentation, and pedestrian, traffic-sign, and vehicle detection.
• Mapping and localisation with camera, lidar, IMU, odometry and GNSS+RTK.
• Planning and decision-making, with PID and Model Predictive Control for actuation.
• Introduced SIL with CARLA and HIL with dSPACE SCALEXIO/IPG CarMaker to validate perception, localisation, planning and control before vehicle integration.
• Delivered the first autonomous shuttle demo with lidar mapping, offline path generation, obstacle detection, emergency braking and mobile-app destination selection.
• Vision-based drone detection and pose estimation from the car module, and autonomous navigation for centring and latching with the drone module, shown at Amsterdam Drone Week 2018.
• Helped turn Italdesign's first software and robotics project into follow-on work and patents.

Gallery

Previous

Next Close

Highlights

• Developed image-based visual servoing for Pepper using ROS and OpenTLD to track and follow generic targets in real time.
• Built a people-following behaviour driving Pepper's base from onboard vision.
• Combined visual servoing with obstacle avoidance for safe indoor navigation among people.
• Delivered a box-grasping demonstration on Pepper built on top of the same perception and visual-servoing stack.

Highlights

• Integrated image-based visual servoing tasks into a QP-based whole-body controller that simultaneously enforced balance, posture and joint-limit constraints.
• Demonstrated whole-body grasping and gaze control on the Romeo humanoid robot, coordinating arm, torso and leg motion from a single visual task.
• Co-authored "Visual Servoing in an Optimization Framework for the Whole-Body Control of Humanoid Robots" published in IEEE RA-L 2017 and presented at ICRA 2017.

Highlights

• Built a shared framework connecting ROS, MATLAB/Simulink and V-REP for fast prototyping of sensor-based robot control algorithms.
• Enabled the same MATLAB/Simulink and C++ control algorithms to be exercised in simulation first and then deployed onto real robots.
• Used across platforms including Romeo, Pioneer P3-DX, Adept Viper s650 and MikroKopter.

Highlights

• Implemented object localisation and single-arm grasping by visual servoing using 2D and RGB-D cameras.
• Extended the approach to coordinated dual-arm manipulation, including a ball-in-maze solving demonstration.
• Developed a door-handle detection and opening task using an RGB-D camera.
• Built a face detection, recognition and face-servoing pipeline for human-robot interaction.
• Combined object detection, model-based tracking, template tracking, 3D point-cloud segmentation, augmented reality and text detection in natural images across demonstrations.
• Authored "Vision-based Manipulation with the Humanoid Robot Romeo" presented at IEEE Humanoids 2016.

Gallery

Previous

Next Close