World-Action Model Architecture: Video Prediction Over Action Imitation

DYNA-2's architecture is built on video generation rather than the Vision-Language-Action (VLA) adapters that have dominated robotics . During pre-training, the model runs a dual objective: it predicts both the next video frame and the next action from the human video corpus . This approach lets the model build an internal model of contact physics and spatial reasoning — learning causal relationships between actions and their physical consequences — that transfers across embodiments without ever seeing a robot during pre-training .

Pre-training corpus: The company built four nested subsets of egocentric human video — 1,000, 10,000, 100,000, and 1,000,000 hours — with no robot action data used during pre-training . After pre-training, the model is fine-tuned on a relatively small amount of real robot data (hours or even minutes) to transfer learned physical intuition to specific hardware .

The First Human-to-Robot Transfer Scaling Law

Dyna Robotics reported what it calls the first cross-embodiment transfer scaling law in manipulation . The finding is significant because it demonstrates that scaling human video data produces predictable, monotonic improvements in robot performance — mirroring the scaling laws that transformed large language models.

• Scaling law on human data: Across all four orders of magnitude (1K → 10K → 100K → 1M hours), all metrics of video-prediction quality improved monotonically, well-described by a power law with no plateau observed .

• Law transfers to unseen robot data: The same checkpoints were scored zero-shot on 39 distinct robot tasks across two stationary bimanual platforms. Performance improved monotonically as the pre-training human video pool grew, before any robot-specific fine-tuning . One analysis showed that holding action-labelled data fixed at 50,000 hours and adding video-only hours dropped zero-shot robot action MSE from 0.340 to 0.120 .

• Post-training performance scaling: After post-training, the mean normalized score across 14 robot tasks rose from 20% → 28% → 45% → 53% of the attainable maximum as pre-training scale increased from 1K to 1M hours .

Task Success Rate Improvements

DYNA-2's performance gains are most visible in three key areas:

• High-precision manufacturing: DYNA-2 raised task success rates from about 20% to 80–90% through increased pre-training scale alone, without any changes to post-training data . This is a 4x to 4.5x improvement purely from watching more human video.

• Zero-shot customer deployments: At real customer sites, DYNA-2 achieved an 87% quality pass rate, compared to 46% for its VLA-based predecessor DYNA-1 — a 41-point improvement under matched post-training budgets . In head-to-head physical evaluations with identical training conditions, DYNA-2 completed tasks than DYNA-1 and won .

• Bottle-cap twisting (13 minutes of data): With only 13 minutes of robot-specific demonstration data, DYNA-2 was adapted to command a pair of five-fingered dexterous hands to twist open a bottle cap . The company's technical report clarifies the figure was roughly of robot data, with pre-training scale lifting success from 10% at smaller budgets to 40–50% at the full 1M-hour budget .

Implications for the Future of Robotics

DYNA-2's results challenge the prevailing assumption that robots require large amounts of task-specific robot data. By demonstrating a scaling law that transfers from human video to robot hardware, Dyna Robotics opens a practical path toward 10 million hours of training data — with the potential to enable robots to master new physical tasks with just hours of local fine-tuning . The model currently operates as a vendor-operated system; Dyna Robotics has announced no public checkpoint, API, or license for DYNA-2 . Independent replication by the robotics research community will be the next step in validating these results .

World-Action Model Architecture: Video Prediction Over Action Imitation

DYNA-2's architecture is built on video generation rather than the Vision-Language-Action (VLA) adapters that have dominated robotics . During pre-training, the model runs a dual objective: it predicts both the next video frame and the next action from the human video corpus . This approach lets the model build an internal model of contact physics and spatial reasoning — learning causal relationships between actions and their physical consequences — that transfers across embodiments without ever seeing a robot during pre-training .

Pre-training corpus: The company built four nested subsets of egocentric human video — 1,000, 10,000, 100,000, and 1,000,000 hours — with no robot action data used during pre-training . After pre-training, the model is fine-tuned on a relatively small amount of real robot data (hours or even minutes) to transfer learned physical intuition to specific hardware .

The First Human-to-Robot Transfer Scaling Law

Dyna Robotics reported what it calls the first cross-embodiment transfer scaling law in manipulation . The finding is significant because it demonstrates that scaling human video data produces predictable, monotonic improvements in robot performance — mirroring the scaling laws that transformed large language models.

• Scaling law on human data: Across all four orders of magnitude (1K → 10K → 100K → 1M hours), all metrics of video-prediction quality improved monotonically, well-described by a power law with no plateau observed .

• Law transfers to unseen robot data: The same checkpoints were scored zero-shot on 39 distinct robot tasks across two stationary bimanual platforms. Performance improved monotonically as the pre-training human video pool grew, before any robot-specific fine-tuning . One analysis showed that holding action-labelled data fixed at 50,000 hours and adding video-only hours dropped zero-shot robot action MSE from 0.340 to 0.120 .

• Post-training performance scaling: After post-training, the mean normalized score across 14 robot tasks rose from 20% → 28% → 45% → 53% of the attainable maximum as pre-training scale increased from 1K to 1M hours .

Task Success Rate Improvements

DYNA-2's performance gains are most visible in three key areas:

• High-precision manufacturing: DYNA-2 raised task success rates from about 20% to 80–90% through increased pre-training scale alone, without any changes to post-training data . This is a 4x to 4.5x improvement purely from watching more human video.

• Zero-shot customer deployments: At real customer sites, DYNA-2 achieved an 87% quality pass rate, compared to 46% for its VLA-based predecessor DYNA-1 — a 41-point improvement under matched post-training budgets . In head-to-head physical evaluations with identical training conditions, DYNA-2 completed tasks than DYNA-1 and won .

• Bottle-cap twisting (13 minutes of data): With only 13 minutes of robot-specific demonstration data, DYNA-2 was adapted to command a pair of five-fingered dexterous hands to twist open a bottle cap . The company's technical report clarifies the figure was roughly of robot data, with pre-training scale lifting success from 10% at smaller budgets to 40–50% at the full 1M-hour budget .

Implications for the Future of Robotics

DYNA-2's results challenge the prevailing assumption that robots require large amounts of task-specific robot data. By demonstrating a scaling law that transfers from human video to robot hardware, Dyna Robotics opens a practical path toward 10 million hours of training data — with the potential to enable robots to master new physical tasks with just hours of local fine-tuning . The model currently operates as a vendor-operated system; Dyna Robotics has announced no public checkpoint, API, or license for DYNA-2 . Independent replication by the robotics research community will be the next step in validating these results .