Dyna Robotics · v2 · 3× · last seen Aug 12, 2026
Dyna-2 is a "World-Action Model" (WAM) for robotics presented by DYNA Robotics on August 10, 2026, that was pretrained for the first time on over 1 million hours of egocentric human video without using robot data. The model jointly predicts next video frames and next actions, enabling the transfer of physical understanding learned from human observation to various robot platforms (stationary arms, humanoid prototypes, dexterous hands) after only a few hours of local Fine-Tuning. According to the company, Dyna-2 increased success rates on precision tasks from 20% to 80–90% and achieved a quality rate of 87% (Zero-Shot) in a customer rollout compared to 46% with its predecessor DYNA-1. The model builds on the already
Features
| Deployment Model | Zero-shot production deployment at new customer sites without additional fine-tuning; predecessor fleet (DYNA-1) already in 24/7 production operation |
| Use Case Scope | Hospitality, logistics, and factory operations (e.g. hotels, restaurants, laundromats, manufacturing) |
| Integrations | Cross-embodiment transfer to bi-manual parallel-jaw arms (incl. 6-DOF YAM arms), semi-humanoid platforms, and dexterous five-finger hands after just hours, or as little as 10–13 minutes of fine-tuning data |
| Platform | World-Action Model (WAM) with dual next-frame and next-action architecture, deployable on stationary robot arms, humanoid prototypes, and dexterous robot hands |
| Release Date | August 10, 2026, Redwood City, California |