Mireille El Gheche

Lead Research Scientist

AI for the physical world

SoftBank Robotics, Zurich, Switzerland

Mireille El Gheche

美礼 · ميراي みれい · Mi-rei

Blog

Talks, interviews, research updates, and a few personal notes.

Zurich Robot Documentary

Interview

A glimpse behind the scenes of the Zurich Robot Documentary. More details coming soon.

Read more ↗

Come for the Robots. Stay for the Future!

Article · November 4, 2026

Reflections from the “Why Zurich Is Built for Robotics” panel on collaboration, keeping talent, and creating more opportunities for women to lead robotics.

Read more ↗

Talking table tennis and robotics on Al Jadeed

Interview · October 30, 2026

An interview on Lebanon’s Al Jadeed channel about my experience with the robot table tennis project and robotics more broadly.

Read more ↗

Learning to Play at Human Speed: From Simulation to Expert-Level Robot Table Tennis

AI & Robotics · October 15, 2026

An EPFL AI Center x Sony AI Impact Talk on Project ACE and bringing robot learning into the real world.

Read more ↗

Selected among 100+ women in AI and robotics in Switzerland

AI & Robotics · September 2026

An honor made even more special by being one of ten women selected for a two-page newsletter interview.

Read more ↗

In conversation with TechDrive Zurich

Interview · August 2026

A conversation about building AI and teaching a robot to play table tennis.

Read more ↗

When intelligence gets physical

AI & Robotics · June 2026

How AI learns to perceive, move, and act in the real world.

Read more ↗

Kuchisabishii (口寂しい)

Japanese & Culture

Kuchisabishii (口寂しい) literally means “lonely mouth.” It describes the mysterious urge to snack when you are not really hungry, but your mouth is simply bored.

Read more ↗

Building smarter AI is easy. Teaching it sarcasm? Yeah, sure.

AI & CultureRead more ↗

What does Mireille mean?

Personal

The meaning behind the name Mireille.

Read more ↗
Mireille El Gheche

Research

My work connects optimization, machine learning, and robotics—from the mathematical foundations to physical AI systems that operate in the real world.

FocusAI & Robotics
ResearchLearning, vision & control
CityZurich, Switzerland

Google Scholar

Research highlights

Featured Publications

Research output

Publications

Mireille El Gheche

Teaching

I have taught university-level courses, supervised PhD students, and mentored engineers and researchers across applied machine learning, AI, and optimization projects. My teaching and mentoring work has focused on helping students and research teams connect theoretical methods to practical problems.


Course in preparation

Language Models

In collaboration with Giovanni Chierchia ↗

A course introducing language models from first principles.

Course introduction

Open on YouTube ↗

Course chapters

  1. 01The idea behind a language model

    In this first chapter, we are going to build the central idea behind a language model from the ground up.

    We will begin with a very simple question: given some text, how can a computer predict what is likely to come next?

    Coming soon

  2. 02The language-model pipeline

    The central message of this lecture is that next-token prediction is the core computational task of a language model, but many other decisions determine its behavior.

    Real applications usually place the model inside a larger system that may include memory, document retrieval, tools, databases, or web services.

    Coming soon

  3. 03From text to tokens

    In this chapter, we focus on the very first part of that pipeline: how text becomes something a language model can process. When we type a sentence into a language model, the model does not directly receive words or characters. The text first has to be converted into numerical representations.

    The first major step in that process is tokenization.

    Coming soon

  4. 04The decoder-only Transformer

    In this chapter, we move to the next part of the pipeline and ask: what does the model actually do with those tokens?

    We will focus on the architecture used by most modern generative language models: the decoder-only Transformer.

    Coming soon

  5. 05How a model learns

    We move to the next question: how does a model actually learn?

    This chapter focuses on pretraining, the stage where a language model develops its broad language capabilities.

    Coming soon

  6. 06Post-training

    In this chapter, we move to the second major training stage: post-training.

    The basic question is: once we have a capable pretrained model, how do we make it behave the way we actually want?

    Coming soon

Mireille El Gheche

About Me

I am a lead research scientist in AI and robotics based in Zurich. My work focuses on taking reinforcement learning, computer vision, sensing, and control from research ideas to integrated physical AI systems that can be tested, demonstrated, and deployed in real environments.

Background

2026 - Present

Lead Research Scientist

SoftBank Robotics, Zurich, Switzerland

2021 - 2026

Staff Research Scientist

Sony AI, Zurich

2017 - 2021

Postdoctoral Researcher

EPFL, LTS4 laboratory

2015 - 2017

Postdoctoral Researcher

Universite de Bordeaux, IMS department

2010 - 2014

PhD in Computer Science

Université Gustave Eiffel (UGE), Paris

Optimization methods for image and signal processing, with applications to computer vision and machine learning.

2009 - 2010

Master of Science in Wireless Communications

CentraleSupélec, Paris

Advanced communication systems and signal processing.

Spoken Languages

ArabicNative
EnglishFluent
FrenchFluent
GermanA2
ItalianA1
Mireille El Gheche

Blog