Guanzhi Wang

Guanzhi Wang is a research scientist at NVIDIA whose research focuses on post-training robot foundation models and agentic robotics. His collaborative work includes Voyager, a Minecraft agent that builds a library of executable skills, and NitroGen, a model that learns game controls from internet gameplay videos.[2][3][7][6]

Wang is a co-lead and co-adviser of NitroGen, with Loïc Magne and Anas Awadalla also credited as co-leads. The model takes pixels as input and predicts gamepad actions, learning through behavior cloning from a gameplay dataset assembled from internet videos. Post-training allows it to adapt to unseen games.[6][7]

The released model has a consequential constraint: its 500-million-parameter diffusion transformer sees only the latest frame. The repository describes it as “a fast-reacting system-1 sensory model” and explicitly says it cannot yet plan over long horizons, play games end-to-end, improve itself over time or play completely unseen games. Its adaptation capability therefore depends on post-training.[7]

Earlier, Wang worked on Minecraft research that supplied a benchmark and internet knowledge corpus through MineDojo, followed by Voyager’s approach to accumulating reusable skills.[4][5][3]

The 2022 MineDojo project credits Wang and Yunfan Jiang as equal contributors and lists them as lead-developer contacts. MineDojo provides a Minecraft multitask benchmark and a corpus of narrated gameplay. Its repository reports more than 730,000 videos, approximately 300,000 hours and 2.2 billion words of English transcripts. Time-aligned narration lets researchers connect natural-language descriptions with video pixels without laborious human labeling; the benchmark includes programmatic, creative and playthrough tasks.[4][5]

Wang is the first-listed author of the 2023 Voyager work, alongside Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan and Anima Anandkumar. In his introduction on the NeurIPS 2023 recap, he identified himself as a third-year Caltech PhD student and NVIDIA research intern presenting Voyager.[3][1]

Voyager combines an automatic curriculum for exploration, an expanding library of executable-code skills and iterative prompting that uses environmental feedback, execution errors and self-verification to improve programs. It queries GPT-4 without fine-tuning the model’s parameters. Storing behaviors as reusable, composable code allows the agent to retain skills and build more complex behaviors from them.[3]

The Voyager authors report that their Minecraft agent collected 3.3 times as many unique items, traveled 2.3 times as far and reached key technology-tree milestones up to 15.3 times faster than prior state-of-the-art methods. These are reported Minecraft comparisons; the milestone figure is a maximum speedup.[3]

Wang’s 2026 Caltech thesis is titled Building Foundation Agents with Internet Knowledge and Large Language Models.[8]