IROS 2022 Tutorial on
Open and Trustworthy Deep Learning for Robotics
Presentation slides: https://drive.google.com/drive/folders/1Tt-tVc6b6TUwdSPJDyb0wOCaoVnBHo-F?usp=sharing
Zoom link: https://us06web.zoom.us/j/87948222129 (Registration is required)
We are happy to announce the half-day tutorial "Open and Trustworthy Deep Learning for Robotics" at the conference IROS 2022. The tutorial will follow IROS style (in-person, online, or hybrid) on October 27rd, from 13:00 to 17:00.
Closer to the event, we will disclose the link to the virtual room, where the tutorial will be held.
Come to join us!
What to expect from this tutorial?
Deep Learning (DL) has dominated artificial intelligence (AI) with a number of spectacular successful applications from self-driving cars to building accurate models that can generate realistic synthetic images and algorithms. Despite many impressive successful applications, integrating DL into robotic systems is still not trivial, since there are many research questions that have not been addressed by the computer vision and machine learning communities yet.
We believe the use of DL within the context of robotics leads to specific learning, reasoning, and embodiment problems that are not appreciated sufficiently by the robotics and computer vision communities. These – along with the steep learning curve and black-box nature of many DL models – often limit the integration of DL-based solutions in robotics, since even models that achieve state-of-the-art performance on standard baselines cannot be directly integrated into real robotic systems. DL also comes with an additional drawback: models are typically very complex, requiring a significant amount of computation power and energy, which can be a limiting factor in many applications, such as mobile robots. Not but the least, the fragmentation of DL development and lack of an open and interoperable ecosystem slows down the integration process, since DL-based solutions require a radically different methodology than conventional robotics methodologies.
DL also suffers from an intrinsic flaw: the inherent complexity and black-box nature of DL models create many safety, trustworthiness, and explainability concerns for their use. Providing explanations regarding the choices and actions of a system is often required by the legislation, while in many safety-critical applications certification is needed in order to ensure that the operation of a robotic system will not put humans or infrastructure at risk. At the same time, the susceptibility of DL models to adversarial attacks further reinforces concerns over the use of DL in safety-critical areas. Indeed, recent literature has demonstrated that even changing one pixel in high-dimensional images can fool models that have not been appropriately trained to withstand such types of attacks. Therefore, the need for trustworthy, safe, and explainable DL is more relevant than ever.
Motivated by the aforementioned challenges, the experts in the tutorial will focus on,
(i) open deep learning tools for robot perception,
(ii) open deep learning tools for robot navigation,
(iii) open deep learning tools for robot control,
(iv) trustworthy deep learning for robotics: form testing approaches to formal methods.
The ultimate goal of this tutorial is to provide a holistic view on the field, bridging the knowledge from several different communities that used to work independently of each other, which is essential to tackle the modern challenges that arise by the rapidly increasing use of deep learning in robotics.
Topics of interest
● Open Deep Learning Tools for Robotics
● Trustworthy Deep Learning for Robotics
● Deep Learning for Robot Perception
● Deep Learning for Robot Navigation
● Deep Learning for Robot Control
● Robot Safety
● Software Architectures for Robotic and Automation
● Robot programming through Imitation Learning
● Robot control optimization through Deep (Reinforcement) Learning
● Object detection and classification with Deep Learning
● Autonomous vehicles applications of DL
● Simulation modeling for Deep Learning training and validation
● Real-world case studies
Confirmed Speakers
Prof. Xiaowei Huang
Prof. Erdal Kayacan
Prof. Anastasios Tefas
Prof. Alexandros Iosifidis
Prof. Abhinav Valada
Prof. Robert Babuska
Prof. Jens Kober
Dr. Nikolaos Passalis