Jason Ma

Jason Ma is co-founder and chief scientist of Dyna Robotics, where he develops models that use predicted future visual states to guide robot actions. His aim is to transfer broad knowledge from human video into robots, then adapt the system to specific physical tasks using robot demonstrations and feedback.[1]

Ma traces his interest in AI to AlphaGo, which surprised him after years of playing Go. At the University of Pennsylvania he pursued a PhD working on reinforcement learning for robotics. He later worked at NVIDIA and Google DeepMind; in late 2024 he founded Dyna with Lindon Gao and York Yang, seeking work that could be used outside the laboratory.[1]

One step in that path was the Eureka project. Ma recounts trying to teach a simulated robot hand to spin a pen and finding it difficult to hand-design the reward function. His team used a language model that could write code to generate reward functions for reinforcement learning. The result showed him a way language models might accelerate robotics research, within the limitations of simulation.[1]

Ma describes Dyna-2 as a world-action model: it takes recent camera images, robot joint information and a language instruction, and produces robot actions alongside predicted future video or a representation of it. The proposed advantage is that the model can use an expected visual future to guide what it does next. This is Ma’s account of the architecture, rather than evidence that prediction alone guarantees reliable control.[1]

He says Dyna-2 was trained on more than one million hours of human data. The breadth of human video gives the model a prior about physical activity, but people and robots have different bodies: Ma says robot data is still needed to bridge that gap and task-specific examples help achieve expertise. For one chopping demonstration, the hosts reported 45 minutes of collected data; Ma describes other tasks taking tens of minutes or a couple of hours, followed by real-world evaluation and iteration.[1]

That principle also shapes the hardware. Ma says Dyna gives its semi-humanoid robot a mobile base and a moving torso, so it can use movements that appear in human training data while remaining practical to deploy. He describes folding napkins for multiple Din Tai Fung locations and towel folding for Best Western; those deployment and volume claims are his account, not independently measured performance.[1]

Ma identifies a possible weakness in vision-heavy training: video does not necessarily capture touch or sound. A model with richer sensing could, in his view, learn some physical tasks with less visual data. He also points to hardware durability and the difficulty of collecting the right data as bottlenecks on more general, longer tasks.[1]