Dyna Robotics Launches Dyna-2 World-Action Model

Dyna Robotics has released Dyna-2, a novel world-action model designed for robot manipulation tasks. This model distinguishes itself by being pre-trained on an extensive dataset comprising over one million hours of egocentric human video footage. This vast amount of data is equivalent to approximately 170 years of continuous human waking experience, a significant leap in the scale of training data for robot learning.

Historically, robot learning has faced a critical bottleneck due to the scarcity of action-labeled data, which typically requires deliberate production through teleoperation. Dyna-2 investigates whether readily available, ordinary human video can serve as a viable substitute for this specialized data. The research team at Dyna Robotics conducted experiments by training a data ladder, progressing from 1,000 to 1,000,000 hours of video, to measure the impact of data scale on model performance. Their findings revealed three key outcomes: the establishment of a scaling law specifically for human video data, the first successful transfer of this law to robot data that the model had not previously encountered, and evidence indicating that video prediction is the primary driver behind this successful transfer.

While Dyna-2 demonstrates significant potential, its current deployment is restricted. It is available as a vendor-operated system rather than through downloadable weights, public checkpoints, APIs, or licenses. This means that to utilize Dyna-2, customers must purchase a Dyna robot cell. The company's prior announcement on August 10, 2026, indicated that Dyna-1 robots are already operational in various service industries, including hotels, restaurants, and laundromats. This suggests that Dyna-2 is targeted towards mid-market service operators and multi-site enterprises that handle repetitive, stationary manipulation work. It is not intended for individual developers or research laboratories seeking local inference capabilities.

The industries identified as suitable for Dyna-2 deployment include hospitality, commercial laundry, food service, light assembly and kitting, and facilities cleaning. The model has been post-trained on 14 distinct tasks that directly map to real-world applications. These tasks include clearing trash trays, assembling first-aid kits, constructing totes, scooping food, tying ropes, preparing hangers, and retrieving specific drinks from refrigerators. Dyna-2 is architecturally a mixture of transformers, with separate tokenization for video and action, utilizing distinct DiT layer stacks that incorporate attention mechanisms to process the complex spatiotemporal data.