Quentin Le Lidec — AMI Labs

Biography

Quentin Le Lidec is a postdoctoral researcher at AMI Labs, where he develops world models for robotics and explores how artificial intelligence can solve problems in the physical world. His research sits at the intersection of machine learning, optimization, and computer vision, with a particular focus on enabling intelligent systems to understand, predict, and act in complex environments.

Before joining AMI Labs, Le Lidec was a postdoctoral researcher at New York University working with Yann LeCun on world models. His recent work includes LeWorldModel, a joint-embedding predictive architecture designed to learn from raw pixels and support efficient planning, as well as stable-worldmodel, an open research platform for reproducible world-model development and evaluation. He also contributes to open-source robotics and simulation tools, including Pinocchio and Simple.

Le Lidec earned his PhD in computer science from École Normale Supérieure, where his research on differentiable optimization for robotic simulation, learning and control was supervised by Justin Carpentier, Cordelia Schmid and Ivan Laptev. His dissertation received the GDR Robotique Best Thesis Award in 2025. He also holds an engineering degree from École Polytechnique and a master’s degree in mathematics, computer vision and machine learning from ENS Paris-Saclay.

Key Papers

LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels

Joint Embedding Predictive Architectures (JEPAs) offer a compelling framework for learning world models in compact latent spaces, yet existing methods remain fragile, relying on complex multi-term losses, exponential moving averages, pre-trained encoders, or auxiliary supervision to avoid representation collapse. In this work, we introduce LeWorldModel (LeWM), the first JEPA that trains stably end-to-end from raw pixels using only two loss terms: a next-embedding prediction loss and a regularizer enforcing Gaussian-distributed latent embeddings. This reduces tunable loss hyperparameters from six to one compared to the only existing end-to-end alternative. With ~15M parameters trainable on a single GPU in a few hours, LeWM plans up to 48x faster than foundation-model-based world models while remaining competitive across diverse 2D and 3D control tasks. Beyond control, we show that LeWM's latent space encodes meaningful physical structure through probing of physical quantities. Surprise evaluation confirms that the model reliably detects physically implausible events.

2026 · arXivLabs

81 citations

More Publications (top 5 by citations)

LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels

2026 · arXiv.org

103 citations

Causal-JEPA: Learning World Models through Object-Level Latent Interventions

2026 · arXiv.org

22 citations

stable-worldmodel: A Platform for Reproducible World Modeling Research and Evaluation

2026 · arXiv.org

2 citations

stable-worldmodel-v1: Reproducible World Modeling Research and Evaluation

2026 · arXiv.org

5 citations

Causal-JEPA: Learning World Models through Object-Level Latent Masking

2026

2 citations