Mail   |   LinkedIn   |   Google Scholar   |   GitHub   |   X   |   CV

Last updated: August 2026

I am a PhD candidate at the Astra-Vision lab, Inria Paris Centre, where I work on video foundation models for physical world understanding.

My research interests lie in making robots and intelligent systems understand and model the real world we live in. I focus on developing AI systems that can perceive, reason about, and interact with physical environments through a deep understanding of the underlying physical principles and dynamics.

We introduce Morpheus, a physics-informed evaluation framework for assessing whether video generative models understand Newtonian dynamics. Morpheus uses 130 real-world videos of physical phenomena and interpretable metrics grounded in conservation laws to measure physical plausibility, showing that today’s models struggle to encode physical principles even when they generate visually convincing videos.

We introduce PIRA (Physics-Informed Representation Alignment), a method for grounding physics into video diffusion models by distilling explicit physical proxy signals — optical flow, depth, segmentation masks, and gravity maps — through the model’s own native 3D-VAE encoder. This yields teacher representations that are inherently compatible with the model’s noisy latents, avoiding the latent-space mismatch of external-teacher distillation. Alignment is performed relationally, matching pairwise token-similarity matrices between student and teacher, and demonstrated on free-fall dynamics with CogVideoX. PIRA is the foundation of an active research thread scaling grounding to multiple interacting objects and richer physical cues.

Source: adapted from Dharmesh Tailor's fork of Leonid Keselman's fork of John Barron's website.