Marina Gallimard

Working on LeRobot, Hugging Face's open-source library for training robot models.

Hugging Face is the reference platform for sharing and developing open-source AI models (over 500,000 models available, used by Google, Meta, Microsoft...).

- Experience with multitasking in Transformer architectures

- Optimization of the training pipeline

- Development of simulation environments to evaluate new models

- In-depth study on validation loss in imitation learning → publication of an official Hugging Face blog post

- Contribution to the development of SmolVLA, their new robotic action generation model

- Co-organization of the robotics workshop for the Women in Data Science conference in Paris

Research work on reconstructing CAD models (3D design files used in engineering) from images, by generating specialized Python code via a multimodal LLM.

- End-to-end ownership of the Reinforcement Learning part of the project

- Design of geometric reward functions: comparison of reconstructed surfaces via nearest neighbor normals in 3D (point cloud)

- Implementation of three asynchronous training variants (Async RL) to speed up training: custom Torch multiprocessing, vLLM generator worker (based on Async RLHF), patching the TRL library for the new point cloud modality

- Experiments on different loss functions (GRPO variants)

- Research on geometric metrics for encoding point clouds and B-Rep (mathematical representation of 3D surfaces)

=> 2 state-of-the-art publications, including an article accepted atCVPR 2026(the world's leading computer vision conference)

I tutored high school students in mathematics for a year.

The work included personalized messaging with each student, preparing additional exercise sheets, and providing guidance for scientific fields.

My approach was: revisit the fundamentals, make abstract reasoning concrete, and provide tools for independent work. Several students were preparing for important exams.