Monday 23rd October, 11:00 - 12:30

Machine Learning for robotics and autonomous vehicles

Slides

Machine learning is at the heart of many functions in robotics and autonomous vehicles, but many challenges remain to be able to apply it easily on real systems. We will provide a quick overview of the use of machine learning, contrast its performances with the human learning capabilities and highlight a few research works trying to reduce this gap, in particular with the goal of improving data efficiency in reinforcement learning.

https://perso.ensta-paris.fr/~filliat/

Monday 23rd, 14:00 - 15:30

Interval analysis with application to the safe navigation of autonomous vehicles

Slides

In this lesson, I will present the basic notions on Interval analysis and constraint networks.
These tools can be used to solve a large class on nonlinear problems such as

1. computing all global minimizers of a non-convex criterion,
2.  computing all solutions of a set of nonlinear equations,
3.  characterizing sets defined by nonlinear inequalities.

Unlike classical numerical approaches (Monte-Carlo or local methods, for instance), the results provided by interval tools are obtained in a guaranteed way, even when strong nonlinearities are involved in the problem.
This property makes interval analysis very attractive to address the safe navigation of autonomous mobile robots.
The purpose of this lesson is to introduce in a pedagogical way the principles of interval methods and contractor programming techniques. To illustrate the efficiency and the reliability of the approach we will consider nonlinear state estimation, dynamic localization and nonlinear control of autonomous vehicles.
Illustrations involving real robots will also be provided.

https://www.ensta-bretagne.fr/jaulin/

Tuesday 24th October, 9:00 - 10:30

Fault Detection Filter Synthesis for a Class of Nonlinear Systems: LMI and Frequency Domain Approaches

This presentation proposes some methods of fault detection filter synthesis for a class of nonlinear systems. First, observer-based LMI synthesis methods for T-S systems subjected to unknown inputs will be presented. Subsequently, multi-objective synthesis problem is discussed in FDI framework. When we are interested in these problems in finite frequency domains (FFD), i.e. in frequency ranges of the fault and unknown perturbations known in advance and belonging to finite frequency bands, these classic techniques (in infinite frequency domains) become quite restrictive. Indeed, the problem of multiobjective synthesis in the finite frequency domain is addressed. In a fault diagnosis context, the generated residue must be as sensitive as possible to faults and as robust as possible against unknown perturbations by means of two finite frequency performance indices such as the H_ and H∞ indexes. Based on the Kalman-Yakubovich-Popov generalized lemma (GKYP) and the Lyapunov method, sufficient design conditions will be presented. Open perspectives will be discussed (such as the use of polyquadratic Lyapunov functions) since the multiobjective LMI synthesis conditions are only sufficient. 

Tuesday 24th, 11:00 - 12:30

Safe trajectory planning for single and multi agents

Slides

Safe trajectory planning consists in, given the present state of a single vehicle or team of vehicles and a map of the environment, computing a trajectory towards a desired goal state or configuration in real-time that optimizes a certain objective function while respecting the kino-dynamic properties of the vehicle(s) and avoiding obstacles and collisions. Even determining the trajectory of a single vehicle can become challenging with the presence of rich obstacles, high-dimensional state space, nonlinear, nonholonomic dynamics, actuation limits, and disturbances. When vehicles can interfere with each other in a common environment that can be partially unknown, the problem becomes increasingly complex. As multi-agent motion planning has a wide range of real-world applications, many approaches have been recently designed to tackle such problems. In the first part of the talk, I will present a short overview of the main classes of methods either for single or multi vehicle safe path planning. The second part of the talk will focus on recent developments where safe path design accounts for the uncertainty on the vehicle motions and the characteristics of the environment. I will also discuss the scalability of the approaches to fleets or swarms of vehicles.

Tuesday 24th, 14:00 - 15:30

Symbolic control of nonlinear systems: safety, optimization and learning

Slides

Symbolic control is a computational approach to control nonlinear systems subject to state and input constraints. It is based on the use of symbolic models that are sound discrete-state approximations of the original dynamics. Symbolic models can be used to automatically synthesize controllers for the original system while providing safety guarantees. In the first part of the talk, I will give a short introduction to the field of symbolic control. The second part of the talk will be devoted to recent developments. I will show how symbolic control and model predictive control can be combined to design controllers with safety guarantees and optimized performance. I will also present recent approaches to compute symbolic models directly from data, paving the way to the development of safe learning approaches for nonlinear systems.

https://sites.google.com/site/antoinesgirard/home

Wednesday 25th October, 9:00 - 10:30

Analysis of visual contents: using complementary information for computer vision applications

Slides

After an introduction of the main concepts in image processing and computer vision, this presentation will focus on methodologies for combination of multi-modal data : 2D images, 3D models, Geometric informations, videos. More precisely, analysis of multi-videos will be presented in the context of crime scene investigation and a method for 2D/3D registration will be described in the context of urban scene analysis.

https://www.irit.fr/~Sylvie.Chambon/

Wednesday 25th October, 11:00 - 12:30

Stepping towards unsupervised 2D/3D scene understanding

Slides

Scene understanding is the ability to extract geometrical and semantic knowledge about a scene by processing sensory data. In this talk, we will deep dive alternatives to the greedy supervised learning setup for either 2D or 3D understanding of complex outdoor scenes. In particular, we will elaborate three axes for relaxing training supervision: multi-task learning with pretext tasks, knowledge distillation from multimodal data (image/text, image/3D), and how simple priors can drive the self-discovery of visual cues.

https://team.inria.fr/rits/membres/raoul-de-charette/

Wednesday 25th October, 14:00 - 15:30

Advanced trends on control and navigation of automated vehicle: human and automated system shared navigation and  context awareness

Autonomous navigation to robotic vehicles has been the focus of many research efforts for several decades. Difficulty robotics issues related to environment perception, mapping and vehicle localization, and the control of displacements were in part solved. However, the functions of automated navigation are used in some specific scenarios and, in most of the cases, as human driver assistance systems. In this way, research has been turned to exploit the shared navigation between the human driver and the onboarded automated system, instead of having a fully automated system. We will consider this topic in this talk, presenting the developments of the Heudiasyc Laboratory in human driver-automated system interaction to improve the navigation of an automated vehicle. The presentation of this symbiotic shared navigation framework will be completed, by considering another advanced trend being developed in Heudiasyc Laboratory, that consists of considering the semantic context of the navigation and ontology modeling in order to improve the vehicle automated navigation.

Thursday 26th October, 9:00 - 10:30

Robustness of neural network? A formal approach to an industrial challenge

Among the numerous properties needed to have a trustworthy AI, in the sense of the future EU regulation on AI, robustness is one of them. This notion has been defined in the ISO/IEC standards as the ability of a system to maintain its performance. While the definition seems rather simple, its application in real life is not that straightforward. One must consider what is the domain of use, what are its characteristics, what kind of perturbation can affect the system etc. All of these parameters have an impact in the way the system can be (and should be) evaluated. This lecture will present the challenges associated with the validation of real world industrialisation of deep learning systems, as well as the way to tackle some of them using formal methods. It will also consider the impact of the standardization works that is currently taking place at the ISO and the EU regulation, on the general best practices that the industry will have to adopt in the following years to come.

Thursday 26th October, 11:00 - 12:30

Safe Reinforcement Learning, Resilient/Fault Tolerant and Health Aware Control Strategies for Autonomous Systems

Slides

Autonomous Systems have gained increased attention during the last years due to their utilization in diverse applications such as surveillance, search and rescue missions, geographic studies, military and security, etc. As almost all process and mission critical systems operate in closed loop, it has become imperative to guarantee the Safety of Autonomous Systems based on the development of control laws that are able to tolerate fault/failure under a reasonable system performance such as well as guarantee desired levels of remaining useful life of the global system. The main aspects will cover new and sophisticated design characteristics and control solutions that satisfy performance demands under faulty conditions or degraded mode of system functioning including reliability and safety requirements.

The talk will present various methodological aspects/strategies that tolerate potential faults/failures to improve reliability, safety, and availability while providing desirable performances. The talk will also present the health aware control systems which can maintain overall system stability and acceptable performance in the event of such faults/failures. Moreover, basic elements of Reinforcement Learning will be considered. Then, the emerging concept of Safe Learning will be presented from both control theory and AI perspectives. The talk will motivate the development of Safe Learning approaches for safety critical systems by taking into account Health-aware/PHM based information. 

Thursday 26th October, 14:00 - 15:30

Intelligent Control Supervisor for Autonomous Vehicles

Slides

Autonomous vehicles, specifically lateral and longitudinal control, are required to perform multiple tasks, such as lane-keeping, lange-change, obstacle avoidance, parking, etc. Such different tasks require different control performances over the whole range of vehicle speeds. Consequently, it is difficult and conservative to design one single controller that can cover all these performances. Thus, it is needed to have a stable and performant control switching architecture containing multiple controllers designed to perform those different tasks.

The following presentation defines the Linear Parameter Varying (LPV) control concept and Youla-Kucera (YK) parameterization, showing their useful uses in autonomous vehicles. In addition, a general switching control architecture is designed using the LPV-YK control concept, analyzing its stability and performance. Then, intelligent switching logic is presented which optimizes the control performance according to the driving situation. The presentation shows real implementation of the proposed approach on a RENAULT ZOE vehicle with results analysis.

Friday 27th October, 9:00 - 10:30

3D LiDAR Localization and Perception for Autonomous Systems

Slides

LiDAR sensors are active sensors able to perceive the environment in the form of a 3D point cloud. They are an important building block of any fully autonomous vehicle and have seen their price dropped over the past ten years. We will see in this talk two Vision tasks on data coming from LiDAR sensors: SLAM (Simultaneous Localization And Mapping) and semantic segmentation of 3D points. For SLAM, we will see that the most efficient methods are not based on neural networks but only on geometry. For semantic segmentation, we will see a method based on geometry and neural networks capable of generalizing to different LiDAR sensors.

Friday 27th October, 11:00 - 12:30

Autonomous Driving Safety Assessment

Part 1: Validation Strategy for Autonomous Vehicles through Simulation 
Part 2: Machine Learning Robustness Evaluation for Autonomous Driving Safety

Demonstrating safety objectives of an automated system such as autonomous vehicles in its traffic environment remains a challenging task with regard to the uncountable number of situations potentially encountered. In the first part of this talk, the idea is to study an approach based on driving scenarios and simulation in order to apply variability to the different parameters of these scenarios and build a massive digital experience plan. We will approach the latest research in safety demonstrations and the problems related to the SOTIF (Safety Of The Intended Functionalities ISO 21448).

In this second half of the talk, we tackle in a more direct way the safety questions arising from the training and validation of Machine Learning based components such as automatic awareness and detection of pedestrians, scene understanding. We will make an overview on how to analyze their robustness towards worst case in-distribution scenarios and their capability of adaptation towards adverse weather conditions such as rain and snow. We will conclude on the emerging use of generative models and their help towards improving the robustness metrics of such ML-based systems.