Jason Ma

Jason Ma is a co-founder of Dyna Robotics, which builds general-purpose robots for useful physical work. His research connects learning a skill with judging whether an action brings a robot closer to its goal. As first author of Eureka and a co-lead author of DrEureka, he helped develop methods that use language models to design training rewards and transfer learned skills from simulation to physical robots.

Learning to recognize progress

Ma completed his PhD at the University of Pennsylvania’s GRASP Laboratory, working with Dinesh Jayaraman and Osbert Bastani. His research experience includes Meta AI, NVIDIA AI and Google DeepMind. Across these collaborations, he explored how human videos and foundation models could supply guidance for robot learning.

In reinforcement learning, a reward is the feedback that distinguishes useful actions from unhelpful ones. Designing that feedback separately for every task can require considerable engineering. Ma’s first-author work on VIP developed visual representations and rewards from video, allowing a pretrained model to guide previously unseen robot tasks. Its implementation was incorporated into PyTorch’s TorchRL library.

In 2023, Ma and collaborators developed LIV, which learned language-image representations and rewards from human videos paired with text. This connected a description of a desired outcome with a visual estimate of progress toward it, bringing perception, task specification and reward learning into one model.

Foundation models as robot teachers

Ma’s subsequent projects expanded the role of foundation models from interpreting activity to designing rewards, training conditions and learning environments.

• Eureka: generating reward code. As first author of Eureka, Ma helped develop a system that uses language models to generate reinforcement-learning reward functions, evaluates the resulting policies and iteratively improves the code. The language model supplies training guidance; reinforcement learning produces the motor skill. A simulated robotic hand spinning a pen illustrated how generated feedback could help a robot learn a dexterous task.

• DrEureka: transferring skills to physical robots. Ma and William Liang shared lead authorship of DrEureka, which automated both reward design and the selection of varied simulation physics used during training. Those variations help prepare a policy for differences between simulation and real hardware. The team demonstrated a physical quadruped balancing and walking on a yoga ball.

• Eurekaverse: generating learning environments. Ma co-advised Eurekaverse, extending language-model code generation to the environments in which robots learn. The system generates increasingly challenging, varied obstacle courses and trains policies against them. Its quadruped parkour experiments explored how the training environment and the learner could develop together.

• Generative Value Learning: estimating task progress. During a Google DeepMind internship, Ma co-led Generative Value Learning with Joey Hejna. The method shuffles video frames and asks a vision-language model to infer their progression toward task completion. This helps distinguish genuine progress from elapsed time. The resulting estimates support success detection, training-data filtering and policy learning without task-specific model training.

Making reliability a product requirement

At Dyna, Ma works alongside co-founders Lindon Gao and York Yang, combining robotics research with their experience building AI-powered hardware businesses. He advocates a research and deployment loop in which customer environments expose weaknesses in models and hardware, supply training data and help determine which research problems deserve attention.

Error recovery is central to this approach. A napkin-folding robot can accidentally pull several napkins from a stack, leaving a scene unlike the successful demonstrations it learned from. Dyna’s reward model estimates progress from robot video. Falling scores help researchers and operators identify troublesome cases, collect targeted recovery demonstrations and fine-tune the policy through human-in-the-loop active learning. Ma also argues that broad training across tasks supplies physical knowledge and recovery behaviors that a robot can reuse in unfamiliar situations.

Dyna’s DYNA-1 experiments reported 99.4% task success during a 24-hour napkin-folding run. Commercial usefulness also depends on fold quality and throughput: completing a fold does not necessarily produce a napkin a restaurant will accept. In its August 2026 deployment update, Dyna described improvements in acceptable output and folding throughput as it expanded deployments with Din Tai Fung.

Ma distinguishes reliable performance at a trained customer site from performance in a new environment. The restaurant napkin-folding and Sacramento laundromat towel-folding deployments he described used site-specific data. He also reported three days of T-shirt folding at CoRL in Korea without additional site data, illustrating the separate challenge of transferring a skill to unfamiliar surroundings.

Dyna’s DYNA-2 world-action model, introduced in August 2026, extends its work on learning from human activity through pretraining on more than a million hours of egocentric video. Ma’s longer-term ambition is a robot platform people can teach and direct across many tasks. His work at Dyna brings that ambition into customer workflows, where learned skills must recover from mistakes and repeatedly meet useful standards.