
Tuesday, November 12, 2024 | 3:00 p.m. - 6:00 p.m.
Politecnico di Milano - Piazza Leonardo Da Vinci 32
Room 3.1.5 (Building 3)
Contacts:
Prof. Elena De Momi | elena.demomi@polimi.it
Prof. Matteo Matteucci | matteo.matteucci@polimi.it
Prof. Andrea Zanchettin | andreamaria.zanchettin@polimi.it
On Tuesday, November 12, 2024, from 3:00 p.m. to 6:00 p.m., the Small Symposium on Robotics, a series of three seminars focused on robotics, will take place in Room 3.1.5 of Building 3 at the Politecnico di Milano. Featured speakers include Prof. Federica Ferraguti (University of Modena and Reggio Emilia), Prof. Gabriele Costante (University of Perugia), and Prof. Andrea Pupa (University of Modena and Reggio Emilia).
Federica Ferraguti – The Future of Surgery: AI-Powered Diagnosis and Robotic Assistance
15:00 – 16:00
Artificial intelligence is about to revolutionize the field of surgery, offering innovative solutions for both diagnosis and procedure assistance. This talk will provide examples of application of AI for automated diagnosis of surgical diseases, exploring how advanced algorithms can analyse medical images and patient data to identify pathologies with high accuracy and precision. Additionally, we will examine the role of robotic assistance in providing support to surgeons during procedures. By integrating AI-powered systems, robotic platforms can offer enhanced precision, dexterity, and visualization, ultimately leading to improved surgical outcomes. A specific use case will be discussed: US-Guided Robotic-Assisted Percutaneous Nephrolithotomy.
Gabriele Costante – Deep Reinforcement Learning in Vision-Based Mobile Robotics: Core Principles, Latest Advances, and Causal Inference Strategies for Data Augmentation
16:00 – 17:00
The development of advanced robots and control systems capable of addressing complex tasks has become a cornerstone in the progression of technology across various fields, including surveillance, agriculture, transportation, and human assistance. Central to this progress are Reinforcement Learning (RL) and Deep Reinforcement Learning (DRL) strategies, which have shown notable effectiveness in modelling perception-to-action relationships. These technologies excel in transforming high-dimensional inputs, such as images, into control commands applicable to a multitude of tasks involving mobile robots.
This seminar will first provide a brief review of DRL principles, followed by a presentation of recent advancements in vision-based DRL methodologies for both terrestrial and aerial robot tasks, including active target tracking, quadrotor control, trajectory tracking, and collision avoidance. The methodological aspects and the experimental results of the proposed solutions will be described, highlighting benefits and limitations. Additionally, the seminar will draw upon these findings to discuss the well-known issue of data efficiency in DRL solutions, whose success is significantly affected by the volume of training data. This is particularly critical when experience collection is conducted using real platforms, as many essential environment interactions, such as those that would result in robot collisions, are restricted or prohibited.
An innovative approach that leverages the causal inference paradigm to establish a counterfactual framework for generating experiences in image-based offline training of DRL policies will be presented. By learning the Structural Causal Model (SCM) that represents the process dynamics, this method generates counterfactual samples for data augmentation.
Andrea Pupa – Planning under Uncertainties: Closed-Loop Sensitivity in Robotic Applications
17:00 – 18:00
Addressing uncertainties is a key challenge in robotics, particularly as robotic systems are increasingly deployed in complex real-world scenarios. Uncertainties in robot dynamics, sensor noise, and environmental conditions can significantly impact planned trajectories, jeopardizing task execution.
This seminar explores recent advancements in robust motion planning, focusing on the concept of closed-loop state sensitivity — a powerful tool for assessing how variations in model parameters impact system behaviour under feedback control. The seminar will cover innovative methods for shaping system trajectories to enhance robustness by minimizing sensitivity. Following this, the concept of uncertainty tubes will be introduced, with an emphasis on their mathematical foundation and effectiveness in constraining deviations from planned trajectories. These approaches have been validated and tested on various robotic platforms, demonstrating their practical applicability and reliability in real-world scenarios.