Maison de la Simulation

Accurate vision-based action recognition is crucial for developing autonomous robots that can operate safely and reliably in complex, real-world environments. In this work, we advance video-based recognition of indoor daily actions for robotic perception by leveraging vision-language models (VLMs) enriched with domain-specific knowledge. We adapt a prompt-learning framework in which class-level textual descriptions of each action are embedded as learnable prompts into a frozen pre-trained VLM backbone. Several strategies for structuring and encoding these textual descriptions are designed and evaluated. Experiments on the ETRI-Activity3D dataset demonstrate that our method, using only RGB video inputs at test time, achieves over 95\% accuracy and outperforms state-of-the-art approaches. These results highlight the effectiveness of knowledge-augmented prompts in enabling robust action recognition with minimal supervision.

In the present work, we have extended molecular Density Functional Theory (MDFT) to study model solvents at high pressure and how chemical reactivity can be modified. Notably, we have considered an example of Diels-Alder reaction in model apolar (CCl4) and polar (CH2Cl2) solvents.MDFT allows to calculate solvation free energies for different chemical structures along the reaction pathway at different pressures. These energies, combined with (electronic) density functional theory calculations providing energetic differences between reactants, transitions states, intermediates and products, allow us to obtain the reaction free energy profiles in a large pressure range (from ambient to 1.5 GPa). Special attention was paid to the role of the solvent dielectric response and its influence on reaction kinetics. The model makes it possible to reproduce the experimental dielectric constant at intermediate pressures (0-0.2 GPa) and to infer its increase at high pressures in the GPa range.The numerical findings are in line with the experimental observations, proving that the reaction is promoted by high pressures and that a trans/cis diastereoselectivity is induced in the product distribution. It is shown that the electrostatic interactions play a major role in these findings.Finally, we can obtain the activation volume, which is a reference quantity in pressure dependentreactivity, as a direct results of our calculations, with values in agreement with what experimentally typically observed.

There is an urgent need for efficient and innovative therapies to treat brain disorders such as psychiatric and neurodegenerative diseases. Immunotherapies have proved to be efficient in many medical areas, but have not been considered to treat brain diseases due to the poor brain penetration of immunoglobulins1,2. Here we developed a bivalent biparatopic antibody, made of two camelid heavy-chain antibodies (called nanobodies)3, one binding to, and the other potentiating the activity of, homodimeric metabotropic glutamate receptor 2. We show that this bivalent nanobody, given peripherally, reaches the brain and corrects cognitive deficits in two preclinical mouse models with endophenotypes resulting from NMDA receptor hypofunction. Notably, these in vivo effects last for at least 7 days after a single intraperitoneal injection and are maintained after subchronic treatment. Our results establish a proof of concept that nanobodies can target brain receptors, and pave the way for nanobody-based therapeutic strategies for the treatment of brain disorders.

We simulate the generation of acoustic and tsunami waves generated by submarine landslides using the linear model developed in [1]. The model is able to reproduce both acoustic and surface gravity waves generated by a moving source (e.g. earthquake, landslide) in a vertically stratified ocean.There are only a few studies that focus on the generation of acoustic waves by submarine landslides. In [2], the combined analysis of field data and simulations underline the presence of an interference pattern in the acoustic waves' spectrogram. The interference pattern has a time-varying bandwidth, which is a signature of the submarine landslide dynamics. In a previous work [3], we used the model developed in [1] to reproduce the interference pattern for a static source. Here we use the same model to simulate a submarine landslide in the 2D case. The simulations reproduce the time-varying bandwidth. They are then used to study the influence of two parameters on the acoustic spectrograms, namely landslide velocity and topography. Different velocity profiles available in the literature [4] are tested. For the topography, we use as reference the 2D case simulated in [1]. We also provide illustrations for the tsunami generation by the landslide. [1] Dubois J, Imperiale S, Mangeney A, Bouchut F, Sainte-Marie J. Acoustic and gravity waves in the ocean: a new derivation of a linear model from the compressible Euler equation. Journal of Fluid Mechanics. 2023;970:A28. doi:10.1017/jfm.2023.595[2] Caplan-Auerbach, J., Dziak, R. P., Bohnenstiehl, D. R., Chadwick, W. W., and Lau, T.-K. (2014), Hydroacoustic investigation of submarine landslides at West Mata volcano, Lau Basin, Geophys. Res. Lett.,  41,  5927– 5934, doi:10.1002/2014GL060964.[3] Dubois, J., Imperiale, S., Mangeney, A., and Sainte-Marie, J.: Simulation of the hydro-acoustic and gravity waves generated by a landslide, EGU General Assembly 2023, Vienna, Austria, 24–28 Apr 2023, EGU23-15937, https://doi.org/10.5194/egusphere-egu23-15937, 2023. [4] Farin, M., Mangeney, A., de Rosny, J., Toussaint, R., & Trinh, P.-T. (2019). Relations between the characteristics of granular column collapses and resultant high-frequency seismic signals. Journal of Geophysical Research: Earth Surface, 124, 2987–3021. https://doi.org/10.1029/2019JF005258

Large deep neural networks (DNNs) have made remarkable progress across various applications. Training a large model requires enormous computational and memory capacity; thus, it requires large-scale parallelization and distribution.