Filip Marić

AI Research Scientist, Meta
San Francisco Bay Area

I am an AI Research Scientist at Meta in the San Francisco Bay Area, where I work on body tracking and generative motion models for AR/VR. My research sits at the intersection of geometric methods and generative deep learning — specifically, how to produce structured, physically-valid motion for articulated bodies. I’m increasingly interested in extending this work toward humanoid robots and embodied AI.

I defended my thesis at the University of Toronto in January 2023, jointly affiliated with the STARS Lab (Jonathan Kelly) and LAMOR at the University of Zagreb (Ivan Petrović). My thesis developed geometric approaches to inverse kinematics and motion planning, drawing on differential geometry and generative models. Before Meta, I spent a year at Samsung Research America in Montreal working on visual-language models for composed image search. I’m happy to hear from people working on related problems.

Selected Projects

XR-Poser: Accurate Egocentric Human Motion Estimation for AR/VR

CVPR 2026 · Meta

A transformer-based model for temporally consistent body pose estimation from egocentric headset cameras, paired with an auto-labeling system that scales training to tens of millions of unlabeled frames. Supports both keypoint and parametric body representations.

Generative Graphical Inverse Kinematics

IEEE T-RO 2024

The first learned generative IK solver that generalizes across robot embodiments. Uses a distance-graph robot representation with a graph neural network to produce diverse solutions in parallel for unseen manipulators.

Riemannian Optimization for Distance-Geometric Inverse Kinematics

IEEE T-RO 2021

A reformulation of inverse kinematics as low-rank Euclidean distance matrix completion, solved on the Riemannian manifold of fixed-rank Gram matrices. Outperforms classical numerical solvers, particularly for multi-end-effector problems.

GraphIK

Open-source library · Python

Distance-graph representations of robotic manipulators for inverse kinematics, motion planning, and learning with graph neural networks. The reference implementation behind several of the IK papers above.

News

Feb 2026
“XR-Poser” accepted at CVPR 2026. preprint
Dec 2024
“Generative Graphical Inverse Kinematics” accepted at IEEE Transactions on Robotics. preprint code
Jun 2024
Joined Meta as an AI Research Scientist, working on body tracking and generative motion models for AR/VR.
Jul 2023
Presented “Euclidean Equivariant Models for Generative Graphical Inverse Kinematics” at the RSS 2023 workshop on Symmetries in Robot Learning. paper
May 2023
Selected for the University of Toronto’s Grads to Watch list.
Feb 2023
Joined Samsung Research America in Montreal, working on visual-language models for composed image search.
Jan 2023
Defended my PhD thesis at the University of Toronto.
Apr 2022
Started a research internship at Meta Reality Labs, working on inside-out body tracking.
Feb 2022
“Convex Iteration for Distance-Geometric Inverse Kinematics” published in IEEE Robotics and Automation Letters. preprint
Oct 2021
“Riemannian Optimization for Distance-Geometric Inverse Kinematics” accepted at IEEE Transactions on Robotics. preprint
Jun 2021
Invited speaker at the RSS 2021 workshop on Geometry and Topology in Robotics. talk video
Oct 2020
“Inverse Kinematics as Low-Rank Euclidean Distance Matrix Completion” won the best workshop contribution award at the IROS 2020 workshop on Bringing Geometric Methods to Robot Learning. paper