Dinghao YANG   杨丁豪

Senior System Engineer, NVIDIA Cosmos Lab

Training infrastructure for Physical AI

Hangzhou, China

Email: dinghaoy@nvidia.com;
             dinghowyang@gmail.com;
Google Scholar: Google Scholar Link
Github: https://github.com/Dinghow
Personal Blog: https://dinghow.site/dinghow-blog

Biography

I am a senior system engineer at NVIDIA, currently working in the Cosmos Lab, where I focus on the development of the Cosmos training framework — with an emphasis on the RL of generative and understanding models. Before NVIDIA, I worked at Lepton AI (acquired by NVIDIA in 2025), where I contributed to the development of Lepton's LLM inference engine. Earlier, at Alibaba Cloud PAI (Platform for AI), I was a core developer of BladeLLM — the official inference engine powering the online services of Qwen.

I received the M.S degree from the School of Electronic and Computer Engineering, Peking University, in 2023, and the B.E. degree from the School of Software Engineering, Tongji University, in 2020.

My interests include GenAI Infra, Reinforcement Learning, Physical AI, and World Foundation Model.

Experience

Publications

Team Publications

Cosmos 3: Omnimodal World Models for Physical AI
NVIDIA, 2026 (I am a core contributor to the Reasoner and Generator RL framework)
[PDF] [Code]
World Simulation With Video Foundation Models for Physical AI
NVIDIA, 2025 (I am a core contributor to the Diffusion RL training framework)
[PDF] [Code]
Cosmos-reason1: From physical common sense to embodied reasoning
NVIDIA, 2025 (I am a core contributor to the SFT & RL training framework)
[PDF] [Code]

First-author Publications

Unified Interactive Image Matting
Dinghao Yang, Bin Wang, Weijia Li, Yiqi Lin, Conghui He
ACM Transactions on Multimedia Computing, Communications and Applications (TOMM), 2025
[PDF] [Code]
PointCHD: A Point Cloud Benchmark for Congenital Heart Disease Classification and Segmentation
Dinghao Yang, Wei Gao
IEEE Journal Of Biomedical & Health Informatics (JBHI), 2024
[PDF] [Code]
Exploring the User Guidance for More Accurate Building Segmentation from High-Resolution Remote Sensing Images
Dinghao Yang, Bin Wang, Conghui He, Weijia Li
International Journal of Applied Earth Observation and Geoinformation (JAG), 2024
[PDF] [Code]
Exploiting Manifold Feature Representation for Efficient Classification of 3D Point Clouds
Dinghao Yang, Wei Gao, Ge Li, Hui Yuan, Junhui Hou, Sam Kwong
ACM Transactions on Multimedia Computing, Communications and Applications (TOMM), 2023
[PDF] [Code]

Projects

Cosmos-RL
Cosmos-RL is a scalable Reinforcement Learning framework for foundation models, one of the first open-source RL frameworks that unifies LLM, VLM, Diffusion, and VLA training under a single abstraction. I lead the infrastructure for the WFM (Diffusion-based World Foundation Model) RL stack, covering training pipeline design and distributed system integration.
[Code]
Lepton Inference Engine
The LLM inference engine developed at Lepton AI (acquired by NVIDIA in 2025). I contributed to model implementation (LLM, low-latency TTS), performance optimization by CUDA kernels, and serving-side performance work.
[Lepton AI]
BladeLLM (Qwen Inference Engine)
BladeLLM is the production inference engine for the Qwen series at Alibaba Cloud. As a core developer, I owned model implementations, serving-side performance work, engine interface development, and Triton kernel optimizations.
[Introduction]
Intelligent Annotation for SenseBee
I am principally responsible for the interactive data annotation algorithms for SenseBee, the data platform for SenseTime AI research. We propose and deploy deep learning-based algorithms to accelerate manual data annotations (i.e. image segmentation, image matting, 3D object detection).
[Code]
OpenPointCloud
I participate in OpenPointCloud, an open-source algorithm library of deep learning-based point cloud compression & processing, mainly contribute to the manifold learning-based point cloud representation learning method.
[Code]

Awards

Academic Activities