Member of Technical Staff · Amazon FAR
I’m building general-purpose robots at Amazon FAR, following the acquisition of Pixelwise AI, where I was a founder. We’re continuing our vision of training robots using internet-scale human data. Previously, I was fortunate to do a PhD at Meta AI and the Visual Geometry Group, University of Oxford, supervised by Christian Rupprecht, Natalia Neverova, and Andrea Vedaldi. Before my PhD, I completed an engineering program at École Polytechnique in beautiful Paris. I also enjoy running and all mountain-related activities 🏔️
Shape reconstruction and tracking from video via agentic analysis-by-synthesis: a VLM agent iteratively refines a structured 3D object model, including its geometry and kinematic structure, through a render-and-compare loop with numerical pose optimisation.
Feed-forward 3D reconstruction scales predictably with model and data size. VGGT-Ω substantially improves accuracy, efficiency and capabilities for both static and dynamic scenes, trained at unprecedented scale with a dynamic-scene annotation pipeline and self-supervised learning.
VGGT is a feed-forward neural net that directly predicts all key 3D attributes of a scene: cameras, point maps, depth maps, and 3D point tracks, from one, a few, or hundreds of its views, within seconds.
In CoTracker3, we propose a simple and effective method for scaling synthetic-trained point trackers on real data.
We propose a new fully differentiable Structure-from-Motion pipeline.
CoTracker bridges the gap between long-term point tracking and Optical Flow by jointly tracking multiple points (pixels) throughout an entire video.
We introduce Dynamic Replica, a synthetic benchmark dataset for dynamic depth-from-stereo models, and propose DynamicStereo, a temporally consistent disparity estimation model that we train on this dataset.