I build generative models to reconstruct, imagine and act in the world.
I am a final-year Ph.D. student at Carnegie Mellon University (CMU)'s Robotics Institute, advised by Jeffrey Ichnowski, and I frequently collaborate with Deva Ramanan and Bowen Wen. I am currently a research intern at World Labs on the pre-training team, under Justin Johnson. Previously, I was a research intern in the DUSt3R Group at NAVER Labs Europe, advised by Jérome Revaud and Vincent Leroy. My interests lie in generative models for spatial intelligence, with a focus on the foundations of generation: diffusion and flow-based modeling, what these models learn, and how pre-training on the physical world transfers to perception and embodied agents.
During my first year at CMU I worked with Sebastian Scherer on geometric camera calibration. Prior to CMU I completed a Bachelor's and Master's degree in Aerospace Engineering at Delft University of Technology, advised by Guido de Croon, studying efficient bio-inspired algorithms for fully autonomous nano drones. In 2019 I was a visiting student at Vijay Janapa Reddi's Edge Computing lab at Harvard University. I am a recipient of the 2023 CMLH Fellowship in Digital Health Innovation and the 2024 CMLH Fellowship in Generative AI in Healthcare.
I am always looking for motivated students to work with.
NeurIPS 2026
PointZero predicts dense 3D point tracks from an RGB-D observation and a few sparse tracks. Pre-training on 2.9M synthetic frames without robot data transfers to real-world dynamics, action-conditioned prediction, and imitation learning.
arXiv preprint, 2026
ModAR generates future point tracks, DINO features, and depth sequentially, then predicts robot actions conditioned on these futures. Training from scratch lets us isolate the effects of modality choice, generation order, and actionless data. The resulting model improves manipulation across simulated and real-world evaluations and benefits from human demonstrations.
World Labs, 2026
Introducing Atlas! An autoregressive and multimodal DiT for generating image and depth frames. Atlas is a state-of-the-art 3D reconstruction and camera-controlled video model trained from scratch on web-scale data. I co-led Atlas-3D with Hao Zhang. We outperform all public models including VGGT-Omega and Pi3X.
arXiv preprint, 2026
Modality Forcing turns a pretrained text-to-image model into a joint image-depth generator with a simple post-training recipe: one DiT, separate noise levels per modality, and per-modality decoders that allow training on sparse real-world depth. Depth accuracy scales with T2I pre-training (300M → 3B), and our strongest model is competitive with state-of-the-art monocular depth estimators, reducing AbsRel by 57% over prior joint image-depth generative models.
arXiv preprint, 2026
3PoinTr learns manipulation from unconstrained human videos: videos where the human demonstrator can act freely rather than mimicking target robot kinematics. 3PoinTr first predicts dense 3D point tracks — how the scene should move to complete the task — and then conditions a closed-loop multitask policy on these tracks. 3PoinTr outperforms strong behavior cloning and learning-from-video baselines across simulated and real-world evaluations.
Conference on Robot Learning (CoRL) 2026
Wiggle and Go! enables zero-shot dynamic rope manipulation through system identification: the robot first wiggles a rope to identify its physical parameters, then uses learned simulation priors to plan an accurate, goal-conditioned throw — without large real-world datasets or iterative retries. The two-stage framework completes dynamic rope tasks accurately on the first attempt across varied ropes and payloads.
NeurIPS 2025
Imagine if robots could fill in the blanks in cluttered scenes. Enter RaySt3R ✨: a single masked RGB-D image in, complete 3D out. It infers depth, object masks, and confidence for novel views, then merges the predictions into a single point cloud.
International Conference on 3D Vision (3DV) 2025, Oral, Best Student Paper Award
MASt3R for SfM with 1000+ unordered images. We contribute a memory-efficient algorithm that leverages the MASt3R encoder for image retrieval without any overhead. MASt3R-SfM has overall linear complexity in the number of images, and handles any set of ordered or unordered images.
Proc. Algorithmic Foundations of Robotics (WAFR) 2024
Deformable objects are common in household, industrial, and healthcare settings; tracking them would unlock applications across robotics, gen-AI, and AR. DeformGS performs dense 3D tracking and dynamic novel-view synthesis on real-world deformable cloths.
project website / arXiv / data / code / X thread
Conference on Robot Learning (CoRL) 2024
Cloth-Splatting accurately estimates the state of deformable objects from RGB supervision. It uses a GNN as a prior to improve tracking accuracy and convergence speed.
International Conference on 3D Vision (3DV) 2025
Online 3D tracking can unlock many new applications in robotics, AR, and VR. Most prior work targets offline tracking on full sequences. DynOMo simultaneously performs 3D tracking, 3D reconstruction, novel-view synthesis, and pose estimation.
IEEE International Conference on Robotics and Automation (ICRA) 2024
🌟 Spotlight 🌟 presentation at the ICCV23 — TRICKY Workshop
Residual-NeRF improves depth perception and training speed for transparent objects. By first learning a background NeRF of the workspace without the transparent objects to be manipulated, we recover better depth quality and faster convergence.
project website / arXiv / video / code
The complete bibliography. The selected works above are repeated here in compact form.
