Humanoid Robotics Technology
Dyna Robotics has announced DYNA-2, its flagship World-Action Model (WAM) and the first robot foundation model trained on over 1 million hours of data, equivalent to 170 years of continuous waking experience.
Powered entirely by human egocentric video, DYNA-2 is based on foundational research that demonstrates the first human-to-robot scaling law for robotics, proving that robot performance improves smoothly and predictably with every added hour of human data, without hitting plateaus. This is a paradigm shift for robotic foundation models; by training on video data alone rather than scarce action data or manual teleoperation, Dyna Robotics has discovered a new scaling axis for physical AI, showing that robotics can bypass its greatest bottleneck, scarce, hand-collected teleoperation data, by using video datasets to predictably scale capabilities.
Dyna Robotics’ robots, powered by its previous DYNA-1 model, are already deployed in production in high-demand environments like hotels, restaurants, and laundromats. DYNA-2 is a major step forward in making production-grade, autonomous robots commercially-viable. In high-precision manufacturing tasks, DYNA-2 increased task success rates from 20% to 80–90% purely through pre-training scale, without requiring any changes to post-training data. DYNA-2 enables zero-shot performance at production-level speeds and quality across new deployment sites.
Action data has long choked generalist robotics; collecting physical teleoperation data manually is expensive, slow, and unscalable. DYNA-2 bypasses this bottleneck by establishing a cross-embodiment transfer scaling law. Pre-trained on 1 million hours of human video using a dual next-frame and next-action world-modeling architecture, DYNA-2 enables physical intuition and spatial reasoning gained from human motion to transfer directly to robot hardware, even without seeing a single frame of robot data in pre-training. This dramatically lowers the barrier to training robots on new tasks.
“For years, generalist robotics has been choked by a data bottleneck: collecting physical teleoperation data manually simply cannot scale to general intelligence,” said Dyna Robotics co-founder Jason Ma. “Action data is scarce, but video is everywhere, and with DYNA-2, we showed that physical intuition doesn’t require millions of hours of training on a robot arm; it can be learned directly from human video. By building a World-Action Model that imagines how the physical world moves before taking action, we give robots spatial reasoning and contact physics that traditional vision-language models simply lack and show that physical AI can scale smoothly and predictably, unlocking commercial-grade automation far faster than previously thought possible.”
Production-Grade Performance in Hours
DYNA-2 can be quickly adapted to perform tasks on many different robot platforms with just a few hours of local fine-tuning:
- Aggregated over 15 benchmark tasks, DYNA-2 policies pre-trained with more human data consistently performed better
- In one case, with just 13 minutes of data, DYNA-2 could command a pair of five-fingered robot hands to twist open a bottle cap
Significant Performance Increases Versus Vision-Language-Action Models
DYNA-2 shifts the foundation model architecture from Vision-Language-Action (VLA) adaptations to a native WAM built on video generation. In head-to-head physical evaluations against DYNA-1, Dyna Robotics’ VLA model, under matched training steps and datasets, DYNA-2 demonstrated decisive control and task resilience advantages:
- In real-world evaluations, DYNA-2 achieved a customer-quality pass rate 1.55× that of DYNA-1, despite both models attaining a 99.9% task success rate. In a zero-shot customer deployment, DYNA-2 achieved an 87% quality pass rate, compared with just 46% for DYNA-1.
- DYNA-2 is able to recover from physical disturbances without human intervention when completing tasks requiring dexterous manipulation like chopping food and clearing workspaces, compared to the VLA baseline which failed in a single loop and required manual intervention to recover.
- DYNA-2’s video co-training algorithm improved scores by 133% on instruction-following tasks where robots must perform distinct motions based on user-provided commands
Source: DYNA Robotics
