INRIA

Landslides and avalanches jointly cause approximately 150 deaths and €4.9 billion economic losses each year, with the impacts predicted to become more severe due to climate change. Mitigation and prevention of disasters requires accurate predictions of these phenomena, which due to their scale is only achievable via modelling and simulation. Accurate models of landslides in permafrost or avalanches must account for micro-scale (<1mm) processes such as cracks and shear bands that also involve thermal and hydrological effects that will be exacerbated by climate change. Such models do not currently exist. Further, this level of refinement is not computationally viable when modelling an entire mountainside, and so a new approach must be adopted. This project will: 1) Develop new models for permafrost and snow subject to climate-change-induced loadings; 2) Use the new data-driven mechanics framework to transfer information from these models to the scale of the mountainside; and 3) Simulate the effects of climate change on the Mont-Blanc massif at Chamonix. This will combine the researcher's experience with shear band models with the supervisor's expertise in crack models and optimisation techniques. A secondment at a group specialising in simulating landslides and avalanches will provide the expertise to implement the simulation on a real mountainside. This interdisciplinary project will ideally set the researcher for a career in academia in Europe, while benefiting the community at Chamonix, in particular the guide's association, as they will be able to plan adaptations and mitigations for the effects of climate change, ensuring their tourism industry remains viable. Specialised multiphysical models that are adapted to permafrost and snow will advance the state-of-the-art significantly, and the implementation of optimisation techniques in data-driven mechanics has wide applicability throughout civil and mechanical engineering, geology and environmental science.

The objective of VHIALab is the development and commercialization of software packages enabling a robot companion to robustly interact with multiple users. VHIALab builds on the scientific findings of ERC VHIA (February 2014 - January 2019). Solving the problems of audio-visual analysis and interaction opens the door to multi-party and multi-modal human-robot interaction (HRI). In contrast to well investigated single-user spoken dialog systems, these problems are extremely challenging because of noise, interferences and reverberation present in far-field acoustic signals, overlap of speech signals from two or more different speakers, visual clutter due to complex situations, people appearing and disappearing over time, speakers turning their faces away from the robot, etc. For these reasons, today's companion robots have extremely limited capacities to naturally interact with a group of people. Current vision and speech technologies only enable single-user face-to-face interaction with a robot, benefitting from recent advances in speech recognition, face recognition, and lip reading based on close-field microphones and cameras facing the user. As a consequence, although companion robots have an enormous commercialization potential, they are not yet available on the consumer market. The goal of VHIALab is to further reduce the gap between VHIA's research activities and the commercialization of companion robots with HRI capabilities. We propose to concentrate onto the problem of audio-visual detection and tracking of several speakers, to develop an associated software platform, to interface this software with a commercially available companion robot, and to demonstrate the project achievements based on challenging practical scenarios.

Spreadsheet applications (such as Microsoft Excel + VBA) are heavily used in a wide range of fields including engineering, finance, management, statistics and health. However, they do not ensure robustness properties, thus spreadsheet errors are hard to avoid, common and potentially very costly. According to estimates, the annual cost of spreadsheet errors is around 7 billion dollars. For instance, in 2013, a series of basic spreadsheet errors at JPMorgan incurred 6 billion dollars trading losses. To avoid such problems, spreadsheet users need better support from rigorous tools, since the development of industrial spreadsheets typically involves multiple tabs, formulas, macros and data read from external sources (e.g., the internet). However, as of today, spreadsheet environment offer almost no verification support, and third party tools do not allow to reason correctly over macros, and future uses of existing formulas. The MemCAD ERC StG project opened the way to novel formal analysis techniques for spreadsheet applications. Indeed, the MemCAD project led to the design of powerful abstraction techniques to reason over complex data-structures such as tables as found in spreadsheets. During this project, we have implemented AiXL, a prototype that was able to analyze large public benchmarks. It uncovered defects that are beyond the scope of other approaches. We propose to leverage these results into a toolbox able to safely verify, optimize and maintain spreadsheets, so as to reduce the likelihood of errors. The envisioned toolbox relies on automatic and conservative semantic static analysis, so that it will report all occurrences of certain classes of errors. It will have an open architecture with plugins, and will easily extend to user specific properties. This works will take place in a partnership with MatrixLEAD, a startup created based on the results of the ERC MemCAD project. MatrixLEAD will provide industrial use-cases and commercialization plans.

Building machines that interact with their world, discover interesting interactions and learn open-ended repertoires of skills is a long-standing goal in AI. This project aims at tackling the limits of current AI systems by building on three families of methods: Bayesian program induction, intrinsically motivated learning and human-machine linguistic interactions. It targets three objectives: 1) building autonomous agents that learn to generate programs to solve problems with occasional human guidance; 2) studying linguistic interactions between humans and machines via web-based experiments (e.g. properties of human guidance, its impact on learning, human subjective evaluations); and 3) scaling the approach to the generation of constructions in Minecraft, guided by real players. The researcher will collaborate with scientific pioneers and experts in the key fields and methods supporting the project. This includes supervisors Joshua Tenenbaum (program synthesis, MIT) and Pierre-Yves Oudeyer (autonomous learning, Inria); diverse collaborators, and an advisory board composed of an entrepreneur and leading scientists in developmental psychology and human-robot interactions. The 3rd objective will be pursued via a secondment with Thomas Wolf (CSO) at HuggingFace, a world-leading company in the open source development of natural language processing methods and their transfer to the industry. By enabling users to participate in the training of artificial agents, the project aims to open research avenues for more interpretable, performant and adaptive AI systems. This will result in scientific (e.g. interactive program synthesis approaches), societal (e.g. democratized AI training) and economic impacts (e.g. adaptive AI assistants). The dissemination, communication and exploitation plans support these objectives by targeting scientific (AI, cognitive science), industrial (video games, smart homes) and larger communities (gamers, software engineers, large public).