First Joint Workshop of Robotics and Mathematics, June 2019
- Dates
- Wednesday 12 June 2019 (10:00-17:30)
This is a multidisciplinary workshop whose purpose is to bring together Robotics and Applied Mathematics at the University of Birmingham. It will open in the morning, with the talks from the Robotics Group led by Professor Rustam Stolkin.
In the afternoon, the many facets of our Applied Mathematics research are going to be presented. As usual, most of the discussion will take place at lunch and coffee breaks, and the dinner will also follow.
Venue
Vaughan Jeffreys Lecture Theatre, School of Education (R19 on campus map (PDF, 430 kB)).
Speakers
Robotics:
- Irum Mehboob (School of Computer Science)
- Maxime Adjigble (School of Computer Science)
- Claudio Zito (School of Computer Science)
Mathematics:
Agenda
10:00 - 10:30 Maxime Adjigble
Title: Human Robot Interaction - Assisted telemanipulation
Abstract: Robotic telemanipulation plays a significant role in accomplishing a variety of challenging tasks in various fields. It enhances human manipulation capabilities in remote and safety-critical environments, where direct human interaction with the scene objects is not considered feasible. For instance, telemanipulation systems are used in the medical field to execute complex surgical procedures and in nuclear settings they help in carrying out complex decommissioning tasks. Despite all the obvious advantages of telemanipulation, the implementation and wide adoption of these technologies is still in a primitive stage due to the nontriviality of using the existing solutions. We will discuss new methods that allows to fully take advantages of telemanipulation systems.
The slides of my talk are available at: Human Robot Interaction - Assisted telemanipulation (PDF, 4.5 MB)
10:30-11:00 Irum Mehboob
Title: What is the “best” optimisation method in deep learning?
Abstract: Deep learning, as a very important part of machine learning, has given advancement in computer vision tasks. It helps in accurate object detection, localisation, classification, robot grasping and manipulation tasks. Deep learning is based on mathematical tools such as back propagation, loss function and optimisation techniques. There are existing optimisation techniques which are used in deep learning to minimise the loss. This means to predict the parameter values which match the ground truth values of the data. Some of them are for small datasets and some are for large datasets like Adam. My goal is to identify new and efficient optimisation technique which can be used both for small datasets and as well as large datasets but with more emphasis on large datasets, which become more and more prominent these days.
11:00-12:00 Claudio Zito
Title: Transferable Learning for Robotic Manipulation Tasks
Abstract: As robot make their way out of factories into human environments, outer space, and beyond, they require the skill to manipulate their environment in multifarious, unforeseeable circumstances. It is clear that the ability of predicting how their actions affect the environment is a key skill that they need to possess to operate in safety and gain our trust in such environments. Humans perform skilful manipulation tasks from an early age on, and are able to transfer behaviours learned on one object to objects of novel sizes, shapes, and physical properties. For robots, achieving those goals is challenging. For one thing, this complexity arises from the fact that physical and geometrical properties of the environment are usually uncertain, or even unknown, but play a significant role for manipulation tasks.
I will present a set of models that can learn transferable skills for manipulation tasks. The key insight is to represent these tasks in terms of local contacts and how these contacts change over time. These models allow us to generalise within and across object categories by exploiting local similarities between objects of different shapes. I will present the mathematical formulation to compute and use such models, focusing on robotic task such as grasping and pushing of novel objects.
You can read more about my research on my personal webpage.
12:00 Lunch (Staff house)
13:00-13:30 Coffee and Tea (Vaughan Jeffreys)
13:30-14:00 Michal Kocvara
Title:Linear and Nonlinear Semidefinite Optimisation by PENNON
Abstract: Michal will give a brief overview of the software package PENNON for large scale linear and for general nonlinear semidefinite optimisation. To answer the question of why we need software for non-linear SDP, he will discuss the Static Output Feedback (SOF) problem formulated first as a problem with bilinear matrix inequalities and then as a problem with polynomial matrix inequalities.
Here are the slides of my talk: Linear and Non-linear Semidefinite Optimisation by PENNON (PDF, 203 kB)
14:00-14:30 Hong Duong
Title: A class of generalised potential games and applications
Abstract: In the seminal work [1], Monderer and Shapley introduced a fundamental concept of potential games. They are a class of games where the incentive of all players to change their strategy can be expressed via a single global function called the potential function. Using the potential function, the existence of pure-strategy Nash equilibria and the convergence to these equilibria have been shown. Potential games have been studied intensively in theoretical research and have been found in a cornucopia of practical applications in economics and other disciplines such as artificial intelligence and computer vision, theoretical computer science, computational social science and sociology and wireless networks. Recently, there has been a growing interest on game theory in machine learning community. However, the definition of potential games is rather restricted because it requires that the simultaneous gradient of the payoff functions is equal to the gradient of the potential function. In this talk, we are interested in multi-player differentiable games. We establish a connection between weighted potential games and symmetrisable matrices and introduce a notion of generalised potential games that is inspired by a newly developed theory on (generalised) gradient flows in the field of partial differential equations. We also discuss some applications and recent interest in game theory in machine learning.
This talk is based on a joint work with H. Dang (Google, Zürich), B. Tang (Graz, Austria) and H. Tran (Esmart Systems, Norway).
[1] D. Monderer and L. S. Shapley. Potential games. Games and economic behavior, 14(1):124-143, 1996.
Here are the slides of my talk: A class of generalised potential games and applications (PDF, 298 kB)
14:30-15:00 Natalia Petrovskaya
Title: Dealing with sparse and noisy data in problems of spatial ecology
Abstract: In the modern world of big data there are still a plenty of ecological applications where spatial data available for analysis are extremely sparse because of financial, labour, and other reasons. In addition, data obtained from field measurements are often noisy. In my talk I will consider several examples taken from spatial ecology where sparse and/or noisy data are used for reconstruction of spatial patterns and evaluation of functionals from spatial data. Among the other results, it will be demonstrated in the talk how using sparse data converts a deterministic problem into the probabilistic one.
Here are the slides of my talk: Dealing with sparse and noisy data in problems of spatial ecology (PDF, 934 kB)
15:00-15:30 Coffee and Tea (Vaughan Jeffreys)
15:30-16:00 Tom Montenegro-Johnson
Title: Microscale swimming robots: propulsion and control
Abstract: Tom will discuss some recent theoretical and experimental advances in the field of artificial swimming particles, focusing on mechanisms of propulsion, fine control, and fabrication, concluding with some novel applications and possible future directions
16:00-16:30 Panayiota Katsamba
Title: Application-driven design of microbots
Abstract: Bio-inspired microswimmers that mimic bacterial locomotion achieve propulsion at the microscale level using magnetically actuated, rotating helical filaments. A promising application of these artificial microswimmers is in non-invasive medicine, for drug delivery to tumours or microsurgery. Panayiota will discuss how mathematical modelling can address two crucial features in the design of microswimmers. First, the ability to selectively control large ensembles and second, the adaptivity to move through complex conduit geometries, such as the constrictions and curves of the tortuous tumour microvasculature.
16:30-17:00 Sam Johnson
Title: How does it feel to be a neural network?
Abstract: Neural networks underpin the most celebrated recent breakthroughs in artificial intelligence. This is a spectacular example of biomimicry: the concept of a network of dynamical elements whose interaction strengths could be tuned to encode information was discovered by observing the brain. I will discuss some recent results on the relationship between digraph topology and dynamics, once again inspired by living systems, and suggest potential synergies with artificial intelligence research.
Here are the slides of my talk: How does it feel to be a neural network? (PDF, 13.3 MB)
17:00-17:30 Coffee and Tea (Vaughan Jeffreys)
17:30+ Dinner
We plan to go for an early dinner. The venue will be announced later, shortly before the workshop. Please inform Sergey whether you are going to join him and the speakers for dinner.