Yuke Zhu

Key Significance

Yuke Zhu is a leading researcher in Robot Learning and Embodied AI.

NVIDIA GEAR Lab Co-Lead: Leads Generalist Embodied Agent research with Jim Fan
GR00T Project Lead: Spearheads NVIDIA’s humanoid robot foundation model development
robosuite Creator: Developed the standard simulation framework for robot learning
Academia-Industry Bridge: Bridges academic research and industry applications through dual roles at UT Austin and NVIDIA

Profile

Item	Details
Position	Associate Professor, UT Austin (2025-), Assistant Professor (2020-2025)
Affiliation	NVIDIA Research Director & Distinguished Research Scientist
Team	GEAR Lab (Generalist Embodied Agent Research) Co-Lead
Lab	RPL Lab (Robot Perception and Learning Lab)
PhD	Stanford University (2015-2019)
Advisors	Fei-Fei Li, Silvio Savarese
Citations	34,000+ (Google Scholar)

Education

Period	Degree	Institution	Notes
2015-2019	Ph.D. Computer Science	Stanford University	Advisors: Fei-Fei Li, Silvio Savarese
2013-2015	M.S. Computer Science	Stanford University
2011-2013	B.S. Computer Science	Simon Fraser University	First Class with Distinction
2009-2013	B.E. Computer Science	Zhejiang University	Dual Degree Program

Career

NVIDIA Research (2020-present)

Director & Distinguished Research Scientist

Year	Work	Impact
2024	GEAR Lab Founded	Co-founded with Jim Fan, generalist embodied agent research
2024	GR00T	Humanoid foundation model announcement
2025	GR00T N1	Open humanoid VLA model release

UT Austin (2020-present)

Associate Professor, Computer Science (2025-) Assistant Professor (2020-2025)

Year	Work	Impact
2020	RPL Lab Founded	Robot Perception and Learning Lab Director
2022	NSF CAREER Award	Robot manipulation research funding
2023	MimicGen	Large-scale data generation from few demonstrations
2024	RoboCasa	Everyday environment simulation framework
2025	IEEE Early Career Award	Recognition for robot learning contributions

Stanford (2013-2019)

Ph.D. Student, Stanford AI Lab

Year	Work	Impact
2017	AI2-THOR	3D indoor environment simulator
2017	Target-driven Navigation	Goal-conditioned visual navigation
2019	Making Sense of Vision and Touch	ICRA Best Paper Award
2020	robosuite	Robot learning simulation framework

Research

Research Areas

Core Domains:
1. Robot Learning - Reinforcement learning, imitation learning
2. Computer Vision - Visual perception, scene understanding
3. Embodied AI - Embodied agents, agent systems
4. Simulation - Robot simulation, synthetic data generation

Research Philosophy

“My goal is to build algorithms and systems for autonomous robots and embodied agents that reason about and interact with the real world.”

Research Evolution

2015-2019 (Stanford): Visual navigation, perception-action loop
2019-2022: Simulation frameworks, robot manipulation
2022-2024: Data generation, foundation models
2024-present: Humanoid robotics, GR00T

Key Publications

Simulation & Benchmarks

robosuite (2020) - Modular simulation framework for robot learning
AI2-THOR (2017) - 3D indoor environment simulator
RoboCasa (2024) - Everyday environment simulation

Robot Learning

MimicGen (CoRL 2023) - Automated large-scale data generation from few demonstrations
DexMimicGen (2024) - Bimanual dexterous manipulation data generation
Making Sense of Vision and Touch (ICRA 2019) - Best Paper Award

Target-driven Visual Navigation (ICRA 2017) - Goal-conditioned visual navigation
Visual Semantic Planning (ICCV 2017) - Semantic planning

Foundation Models

GR00T N1 (2025) - Humanoid robot foundation model
MineDojo (NeurIPS 2022 Outstanding Paper) - Minecraft-based agent benchmark

GR00T Project

Role

Yuke Zhu is a key lead of NVIDIA’s GR00T (Generalist Robot 00 Technology) project.

GR00T N1 Architecture:
- Dual-system design (System 1 + System 2)
- Vision-Language Module (System 2): Environment interpretation, language understanding
- Diffusion Transformer (System 1): Real-time motor action generation
- Support for various humanoid robots

Key Contributions

Architecture Design: VLA (Vision-Language-Action) model structure
Data Pyramid: Leveraging full spectrum from real to synthetic data
Open Source Release: Democratizing research through GR00T N1 release
Simulation Integration: Integration with Isaac Lab, robosuite

GR00T N1 (2025)

NVIDIA’s first open humanoid foundation model
Natural language instruction understanding and execution
Human motion imitation learning
Support for various robot embodiments

GEAR Lab

Collaboration with Jim Fan

GEAR Lab is an NVIDIA Research group co-led by Jim Fan and Yuke Zhu.

GEAR Lab Research Areas:
1. LLM for Planning - Large language model-based planning
2. Vision-Language Models - Vision-language models
3. Robotic Systems - Robot systems, manipulation, locomotion
4. Simulation Infrastructure - Simulation infrastructure, synthetic data

Division of Expertise

Researcher	Strengths	Representative Projects
Jim Fan	LLM, game agents, communication	Voyager, Eureka
Yuke Zhu	Robot systems, simulation, manipulation	robosuite, MimicGen

Joint Projects

GR00T / GR00T N1: Humanoid foundation model
MineDojo: Minecraft agent benchmark (NeurIPS 2022 Outstanding Paper)
GEAR Research Infrastructure: Isaac Lab, Omniverse integration

Awards & Honors

Major Awards

Year	Award	Organization
2025	IEEE RAS Early Career Award	IEEE
2022	NeurIPS Outstanding Paper Award	NeurIPS (MineDojo)
2022	NSF CAREER Award	National Science Foundation
2022	Outstanding Learning Paper Award	ICRA
2019	Best Conference Paper Award	ICRA
2019	Best Cognitive Robotics Paper (Finalist)	IROS
2021	Amazon Research Award	Amazon
2021	Best Multi-Robotic Systems Paper (Finalist)	ICRA

Corporate Research Support

Amazon Research Award (2021)
JP Morgan Faculty Award
Sony Research Award

Open Source Contributions

robosuite

robosuite: Modular simulation framework for robot learning

Features:
- MuJoCo physics engine based
- 10 commercial robot models supported (including GR1 humanoid)
- 9 grippers, 4 bases supported
- Photo-realistic rendering
- Jointly maintained by Stanford SVL, UT RPL, NVIDIA GEAR

Impact:
- Standard simulation environment for robot learning research
- Used in thousands of research projects

MimicGen

MimicGen: Automated large-scale learning data generation from few demonstrations

Results:
- Less than 200 human demonstrations → 50,000+ auto-generated trajectories
- 18 tasks, multiple simulator support
- Validated on real robots

Impact:
- Dramatically reduced data collection costs
- Enables large-scale data acquisition essential for foundation model training