Jason Ma

Jason Ma was identified in his 2026 AI Engineer World’s Fair biography as co-founder and CTO of Dyna Robotics, a company building general-purpose robots powered by embodied AI foundation models. Previously, he was a research scientist at DeepMind focused on foundation models for robotics.

Ma presents approaches to making generalist robot policies reliable enough for commercial work. He describes Dyna’s research and deployment cycle, in which customer deployments supply training data and expose weaknesses that guide research. The training approach combines off-robot, robot-task and deployment data with an architecture pairing high-level reasoning and a low-level world action model. Reward models score task progress from robot video, flagging mistakes so humans can collect targeted recovery demonstrations and fine-tune the policies through active learning.

In his presentation, Ma reports that Dyna-1 achieved 99.4% napkin-folding success over 24 hours, including recovery from manipulation errors while meeting restaurant fold-quality requirements. He also describes restaurant napkin-folding and Sacramento laundromat towel-folding deployments that used site-specific data, alongside three days of T-shirt folding at a conference in Korea without additional site data. His account emphasizes how broad pre-training supplies physical knowledge and recovery behaviors that task-specific training alone can miss.