Research Scientist Jiageng Mao Announces Completion of PhD at USC and Joining Google DeepMind
Research scientist Jiageng Mao announced that he has just completed his PhD at the University of Southern California (USC) and has joined Google DeepMind to work on 3D encoding for Gemini. The post did not specify his position, team size, or launch date.
His public resume states that he started his PhD in Computer Science at USC in September 2023, under the supervision of Yue Wang, focusing on physical artificial intelligence, covering robotics, computer vision, and language models, with an emphasis on visual language action models and world models. His PhD is expected to last until October 2026. He previously worked as a student researcher at DeepMind from June 2025 to January 2026 and has also been at NVIDIA Research.
His funding records include the 2025 NVIDIA Graduate Fellowship and the 2024 Qualcomm Innovation Fellowship. USC nominated him for the Apple AI Scholar and Google PhD Fellowship. In 2026, he will be included in Forbes' 30 Under 30 Asia list.
On the publication side, PhysBench was selected for an oral presentation at ICLR 2025, approximately 1.8% of submissions. Humanoid Everyday won Best Paper at the CVPR Embodied Intelligence Workshop, and Ψ₀ won Best Paper at the CVPR 3D Large Language and Visual Language Action Workshop. In 2019, he also had an oral paper on point cloud interpolation convolution at ICCV. He completed his master's in Electronic Engineering at The Chinese University of Hong Kong from 2018 to 2022.
This position integrates 3D encoding into Gemini rather than continuing to publish independent robotics papers. The transition from student researcher to formal research scientist shifts the focus from physical artificial intelligence to 3D representations in model harvesting. He did not specify whether the encoding will first be used for robotics, video, or spatial understanding.
This is a talent movement, not a model release. The buyer is DeepMind, which aims to enhance Gemini's 3D capabilities, while the seller is a researcher who has just completed a PhD in physical artificial intelligence. The event is driven by the onboarding post. Funding is not disclosed, but research time has shifted from university projects to the Gemini route. The beneficiaries are model teams that require spatial representations, while the pressure is on laboratories still relying on scholarships in the same direction.
Source: Public Information
ABAB AI Insight
Jiageng Mao's path involves point clouds, embodied data, and world models, rather than dialogue fine-tuning. His 2019 ICCV oral paper focused on point cloud convolution, and in 2025, PhysBench was accepted for an oral presentation at ICLR, while in 2026, Humanoid Everyday and Ψ₀ won Best Paper at related CVPR workshops. During his three years at USC for his PhD, he has already been a student researcher at DeepMind and has spent time at NVIDIA Research. His formal employment simply transitions the same direction from visiting to a permanent position.
Capital has appeared on both sides. The NVIDIA Graduate Fellowship and Qualcomm Innovation Fellowship funded him during his PhD, while DeepMind is now funding his full-time role in 3D encoding for Gemini. Being listed in Forbes' 30 Under 30 Asia has transformed him from a paper author into a name that companies can recruit. As a result, his research time has shifted away from his advisor Yue Wang's topics and into the product route of Gemini. The 3D encoding has not received separate funding; it is absorbed into the model budget.
A comparison can be made to computer vision PhDs in the 2010s transitioning from laboratories to Google Photos and Waymo, rather than large language model PhDs only focusing on text alignment. Physical artificial intelligence is currently in a phase of being absorbed by model companies: robotics papers are still winning awards at conferences, while formal positions are now being written as 3D modules within Gemini.
Structurally, this represents a reconstruction of the industry chain. 3D representations, which were previously limited to point cloud papers and robotic data, are now set to become a coding layer in general models. The mechanism involves student researchers first validating directions, and upon completing their PhDs, transitioning into research scientists. Laboratories retain the papers, while model companies keep individuals who can contribute to Gemini.
ABAB News · Cognitive Law
Once student researchers complete their work, formal positions take over the direction.
The next stop for 3D papers is a layer within general models.
Scholarships retain talent, while product routes ultimately determine pricing.