Staff Research Scientist · Google DeepMind
Video generationControllabilityPersonalisationPhysical reasoning
I am a researcher at Google DeepMind, working on video generative models in GenAI. I work on pre-training for Gemini Omni, a frontier multimodal model released at Google I/O 2026. I focus on new capabilities and controls like reference images, audio, and editing.
Before Omni, I worked on Veo, including ingredient-to-video and Veo 3.1, building controllable video systems for character consistency, creative production, robotics evaluation, and physical world understanding.
I finished my PhD at the University of Toronto with Prof. Quaid Morris and David Duvenaud. I did multiple internships at Google Brain, including projects on discrete-variable RL for protein design and similarity networks for cancer prediction.
Controllability, physical realism and runtime efficiency are crucial next steps for video world models. A leap in controllability is required to turn these models into reliable reasoning tools for real-world applications like robotics and will drive wider adoption of these models.
Google I/O 2026
A native multimodal generation model for creating and editing video from any multimodal inputs like images, video, audio or text.
On Gemini Omni, I worked on large-scale pre-training for new reference-driven capabilities, including image and audio conditioning, spanning data generation, training, inference, and evaluation.
Google I/O 2025
Ingredient-driven video generation that preserves identity, motion, and intent.
I was a core contributor to building Veo Ingredients from the ground up, enabling videos to be conditioned on ingredient images of characters, scenes, objects, textures, lighting, style, and more. I then co-led bringing this capability to Veo 3 and Veo 3.1, where the challenge was extending character consistency to videos with sound, including speaking characters, for the first time.
Shows Veo can become a scalable robotics evaluation environment, predicting policy rankings, out-of-distribution failures, and safety issues across 1600+ real-world trials.
Gemini Robotics Team, Coline Devin, Yilun Du, Debidatta Dwibedi, Ruiqi Gao, Abhishek Jindal, Thomas Kipf, Sean Kirmani, Fangchen Liu, Anirudha Majumdar, Andrew Marmon, Carolina Parada, Yulia Rubanova, Dhruv Shah, Vikas Sindhwani, Jie Tan, Fei Xia, Ted Xiao, Sherry Yang, Wenhao Yu, Allan Zhou
arXiv preprint, December 2025 PaperProject page
Advances general-purpose robots with motion transfer across embodiments and multi-level reasoning, connecting world understanding to executable action.
Gemini Robotics Team
arXiv preprint, October 2025 Paper
Gives video teams a motion-sensitive evaluation tool that detects temporal failures missed by FVD and localizes where generated videos go wrong.
Kelsey Allen, Carl Doersch, Guangyao Zhou, Mohammed Suhail, Danny Driess, Ignacio Rocco, Yulia Rubanova, Thomas Kipf, Mehdi S. M. Sajjadi, Kevin Murphy, Joao Carreira, Sjoerd van Steenkiste
arXiv preprint, April 2025 PaperProject page
Shows how diffusion models can control individual objects by appearance and 3D pose, moving image generation toward editable scene-level assets.
Ziyi Wu, Yulia Rubanova, Rishabh Kabra, Drew A. Hudson, Igor Gilitschenski, Yusuf Aytar, Sjoerd van Steenkiste, Kelsey R. Allen, Thomas Kipf
NeurIPS 2024 PaperProject page
Scales learned rigid-body simulation to hundreds of objects and detailed shapes, making neural physics more plausible for real scenes.
Yulia Rubanova, Tatiana Lopez-Guevara, Kelsey R. Allen, William F. Whitney, Kimberly Stachenfeld, Tobias Pfaff
NeurIPS 2024 PaperProject page
Learns editable 3D dynamics directly from observation, bridging video understanding and simulation without privileged physical state.
William F. Whitney, Tatiana Lopez-Guevara, Tobias Pfaff, Yulia Rubanova, Thomas Kipf, Kimberly Stachenfeld, Kelsey R. Allen
ICLR 2024 PaperProject page
Makes learned simulators significantly more accurate and efficient for rigid contact, a hard case for neural physical reasoning.
Kelsey R. Allen, Yulia Rubanova, Tatiana Lopez-Guevara, William F. Whitney, Alvaro Sanchez-Gonzalez, Peter Battaglia, Tobias Pfaff
ICLR 2023 PaperProject page
Uses learned constraints and test-time optimization to combine neural simulation with ideas from traditional physics solvers.
Yulia Rubanova, Alvaro Sanchez-Gonzalez, Tobias Pfaff, Peter Battaglia
ICML 2022 PaperProject page
Extends continuous-time neural dynamics to irregularly sampled time series and latent-variable models.
Yulia Rubanova, Ricky T. Q. Chen, David K. Duvenaud
NeurIPS Best Paper Award winner that introduced continuous-depth neural models and helped establish differentiable dynamics as a foundation for modern ML systems.
Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, David Duvenaud