Dyna Robotics trains DYNA-2 on more than 1 million hours of human video
The founders behind Caper AI and a former DeepMind researcher are betting that human video can reduce the robot data needed for new tasks, though the results remain Dyna Robotics-reported.
By Ryan Merket · Published
Why it matters
Dyna Robotics is trying to reduce reliance on costly task-specific robot demonstrations by pretraining on abundant human video. If Dyna Robotics' reported results hold up under independent testing, robotics teams could adapt machines to new work with less local data.
Dyna Robotics founders Lindon Gao, York Yang and Jason Ma introduced DYNA-2 on August 10, 2026, pitching a robot foundation model trained on more than 1 million hours of first-person human video.
The three founders came to that bet from different sides of physical AI. Gao and Yang previously built Caper AI, which put cameras, sensors and checkout software into grocery carts. Instacart acquired Caper AI for $350 million in October 2021. Ma completed a PhD at the University of Pennsylvania's GRASP robotics lab and worked on robot foundation models and reinforcement learning after research stints at Google DeepMind, NVIDIA AI and Meta AI.
That mix explains Dyna Robotics' approach: build a research model around data that can scale, while testing it on mundane commercial work where failures are measurable. Dyna Robotics says systems powered by its earlier DYNA-1 model have worked in hotels, restaurants and laundromats.
What "trained on video" means
Dyna Robotics calls DYNA-2 a World-Action Model, or WAM. Dyna Robotics' announcement describes a video-based system that predicts future frames and robot actions, with human video supplying the pre-training signal.
Ma summarized the thesis in the announcement: "Action data is scarce, but video is everywhere."
Dyna Robotics says DYNA-2 was pretrained on more than 1 million hours of egocentric human video without seeing robot frames during pre-training. Dyna Robotics also reported improvement across 15 benchmark tasks as the amount of human-video pre-training increased.
Dyna Robotics says its results show a human-to-robot scaling relationship.
From scaling claims to physical tasks
Dyna Robotics reported that success rates on high-precision manufacturing tasks increased from 20% to 80%-90% as pre-training increased, without changes to post-training data. These are Dyna Robotics' internal results.
In matched real-world evaluations, Dyna Robotics reported that DYNA-2 completed tasks 1.55 times more often than DYNA-1, Dyna Robotics' earlier vision-language-action model. At one customer deployment, Dyna Robotics reported an 87% pass rate for DYNA-2 and 46% for DYNA-1.
Dyna Robotics' deployment figures offer a sharper commercial test, but they remain internally reported. A robot can complete an evaluation and still be too slow, inconsistent or expensive for paid work. Independent testing and fuller deployment data would be needed to assess how the pass rates translate across customers and hardware.
A second-time founder bet backed with $143.5 million
DYNA-2 arrives about two years after Dyna Robotics was founded in 2024. Gao's route to robotics ran through finance and repeated attempts at physical retail technology. A Stuyvesant Alumni Association profile says he studied business at NYU Stern, worked in investment banking at Goldman Sachs and built QueueHop, an anti-theft apparel-tag business accepted into Y Combinator's Winter 2016 batch, before developing Caper AI.
Yang was Caper AI's chief technology officer and later worked as a principal research engineer at Instacart. Ma brought experience as a former DeepMind research scientist focused on robotics foundation models. DYNA-2 turns those founders' combined retail-hardware and robotics backgrounds into Dyna Robotics' core product strategy.
Investors have financed the effort heavily. Dyna Robotics raised a $23.5 million seed round co-led by CRV and First Round Capital in March 2025. A $120 million Series A followed in 2025, led by RoboStrategy, CRV and First Round Capital. Salesforce Ventures also said it participated.
That gives Dyna Robotics at least $143.5 million in disclosed financing. The capital buys the compute, data operations, hardware work and field deployments required to test whether DYNA-2's performance keeps improving beyond 1 million hours.
Competitors are also developing models intended to work across tasks or machines. Physical Intelligence describes openpi as an open-source collection of robotics models and packages that includes pi0, a vision-language-action model. Skild AI describes its Skild Brain as shared intelligence designed to control different robot types and tasks. Google DeepMind says Gemini Robotics On-Device is optimized for local, low-latency inference and can adapt to new tasks through fine-tuning. Dyna Robotics is making a more specific wager: human activity video can become a compounding input for dexterous manipulation with less reliance on robot teleoperation data.
Customer economics will decide whether that technical advantage lasts. Dyna Robotics' customer-site pass-rate comparison is encouraging internal evidence. Fleet size, deployment time, maintenance burden and the amount of engineering required for each new workflow will determine whether DYNA-2 becomes a repeatable product or an expensive research lead.