Jason Ma

Hi there! I am co-founder and lead research at Dyna, where we are building full-stack, general AI robots. We are pushing towards a world of physical abundance by combining frontier robotics research, hardware, and deployments into a single product that can be used by millions. If you are interested in our work and mission, please reach out!

Recent highlights from our team:

Dyna-2.1: A Physical Agent for End-to-End Workflows

Introducing Taku, our semi-humanoid robot, and a learning system that leverages human experience to take over entire workflows without frequent human intervention.

Dyna-2: A 1-Million-Hour Scaling Law for WAMs

A world-action model pre-trained on over one million hours of egocentric human video, exhibiting the first human-to-robot transfer scaling law.

Not Just a Model, But a Product

Dyna has scaled deployments across multiple tasks and customers, from restaurants and hotels to logistics and data centers. Din Tai Fung is now expanding Dyna's robots to all of its locations and commissaries.

Selected recent talks on our research and deployment progress:

Building Embodied AGI Through High-Performance Robots That Scale (Fellows Forum, Sep 2026)

Dyna-2: A One Million Hour Scaling Law for World-Action Models (Actuate, Aug 2026)

Google Scholar Github LinkedIn Twitter

yechengma at gmail dot com

The rest of this page is no longer maintained as of 2026.

Prior to Dyna, I was a researcher at Google DeepMind, NVIDIA AI, and Meta AI. I completed my PhD at Penn, focusing on large-scale reinforcement learning methods for robot learning. Even earlier, I studied CS and math at Harvard.

Selected honors:

Selected recent talks:

My research has pioneered new algorithms for training and leveraging foundation models from internet data to teach robots new tasks. Specifically, my research has

• introduced a new paradigm of scaling large language models (LLMs) test-time compute to automate sim-to-real robot learning (ICLR 2024; NVIDIA Top 10 Research Project of Year 2023, RSS 2024, CORL 2024 Oral)
• developed scalable and principled reinforcement learning algorithms that leverage human videos to learn visual rewards for self-improving robotic manipulation (ICLR 2023 Spotlight, ICRA 2024 Best Paper Finalist in Robot Vision; CORL 2023 LEAP Workshop Best Paper, ICRA 2025)
• demonstrated how to best combine multi-modal learning and RL to train and utilize vision-language models (VLMs) for universal value functions (ICML 2023, ICLR 2025 Spotlight)

Many of these methods have been incorporated into production-level simulation and deep learning software, such as NVIDIA Isaac Sim and Pytorch's official reinforcement learning library TorchRL.

Some robots I have trained in my research:

Selected publications as first or last author; full list on Google Scholar.

(* indicates equal contribution, † indicates equal advising)

Vision-Language Models are In-Context Value Learners

Jason Ma*, Joey Hejna, Ayzaan Wahid, Chuyuan Fu, Dhruv Shah, Jacky Liang, Zhuo Xu, Sean Kirmani, Peng Xu, Danny Driess, Ted Xiao, Jonathan Tompson, Osbert Bastani, Dinesh Jayaraman, Wenhao Yu, Tingnan Zhang, Dorsa Sadigh, Fei Xia
International Conference on Learning Representations (ICLR), 2025 (Spotlight)
Webpage • Arxiv •

Articulate-Anything: Automatic Modeling of Articulated Objects via Vision-Language Models

Long Le, Jason Xie, William Liang, Hung-Ju Wang, Yue Yang, Jason Ma, Kyle Vedder, Arjun Krishna, Dinesh Jayaraman, Eric Eaton
International Conference on Learning Representations (ICLR), 2025
Webpage • Arxiv • Code

Eurekaverse: Environment Curriculum Generation via Large Language Models

William Liang, Sam Wang, Hungju Wang, Osbert Bastani, Dinesh Jayaraman†, Jason Ma†
Conference on Robot Learning (CoRL) (Oral) , 2024
Webpage • Arxiv • Code

On-Robot Reinforcement Learning with Goal-Contrastive Rewards

Ondrej Biza, Thomas Weng, Lingfeng Sun, Karl Schmeckpeper, Tarik Kelestemur, Jason Ma†, Robert Platt†, Jan-Willem van de Meent†, Lawson L. S. Wong†
International Conference on Robotics and Automation (ICRA), 2025
Arxiv •

DrEureka: Language Model Guided Sim-To-Real Transfer

Jason Ma*, William Liang*, Hungju Wang, Sam Wang, Yuke Zhu, Linxi "Jim" Fan, Osbert Bastani, Dinesh Jayaraman
Robotics: Science and Systems (RSS), 2024
Webpage • Arxiv • Code

Eureka: Human-Level Reward Design via Coding Large Language Models

Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi "Jim" Fan†, Anima Anandkumar†
International Conference on Learning Representations (ICLR), 2024
NVIDIA Top 10 Research Projects of 2023
Webpage • Arxiv • Code

Universal Visual Decomposer: Long-Horizon Manipulation Made Easy

Charles Zhang*, Yunshuang Li*, Osbert Bastani, Abhishek Gupta, Dinesh Jayaraman, Jason Ma†, Lucas Weihs†
International Conference on Robotics and Automation (ICRA) (Best Paper Finalist) , 2024
Best Paper Award, CORL 2023 LEAP Workshop
Webpage • Arxiv • Code