I have been following a robotics company for a long time, Dyna， When he first started his business in 2024, he felt that his approach was very pragmatic and did not follow the path centered around high-end humanoid robots. The idea at that time was to replace the repetitive labor scenarios of some offline entities in the United States. In just two years, they have indeed accumulated on this path and should be one of the first robots in the market to enter commercial deployment scenarios. The latest release is Dyna-2.1, a long-term autonomous end-to-end workflow physical agent. The founder is a Chinese who started a new business with an old team and has succeeded once before. Lindon Gao：CEO。 Previously co founded Caper AI with York Yang, which was acquired by Instacart for approximately $350 million in 2021; Afterwards, he served as Vice President of Engineering at Instacart and came out to work at Dyna in 2024. York Yang: CTO, co-founder/project leader of Caper. Jason Ma: Chief Scientist. Dr. Bin, formerly a researcher at DeepMind, NVIDIA, Meta AI, and the main author of reinforcement learning papers for robots such as Eureka, responsible for basic models and learning algorithms. -The company's strategy is to first achieve commercial level reliability in one thing, and then expand to the workflow. The team includes research and engineering personnel with backgrounds in DeepMind/OpenAI/Google, as well as hardware implementation capabilities. In the early days, dual arm workstations (fixed or wheeled bases) were first developed, and in 2026, a semi humanoid Taku with wheels will be introduced (taken from the Japanese word for "craftsman" Takumi) -Compared to robots that previously completed some tasks in real-life scenarios, Dyna-2.1 aims to rotate real employees and complete shift handover work in scenarios such as dry cleaning shops, without the need for human intervention throughout the entire process -The business model leans towards Robots as a Service: charging monthly fees based on robots, including hardware, operations, and model updates. Multiple shifts have been deployed in hotels, restaurants, laundry shops, gyms, and other locations. -The learnability of the model: Each skill is based on the most cost-effective teaching data. In a fixed folding deployment, a new service can now reach production standards in just three days, whereas previously it required weeks or even months of on-site engineering -Model Evolution: 1) DYNA-1 (April 2025): A fundamental manipulation model for business operations. Public demonstration: Over 700-900 napkins are stacked without human intervention within 24 hours, with a success rate of approximately 99.4% and a speed of about 60% of a human. Use reward models to estimate progress and correct errors midway. 2) Dyna-2/WAM (World Action Model, August 2026): A pre trained world action model (video diffusion backbone, simultaneously predicting future videos and action blocks) on millions of hours of human first person videos. Emphasize the scaling law of human video to robot migration; The zero sample pass rate at the customer site is approximately 87% (compared to approximately 46% for DYNA-1). 3) Dyna-2.1: Add an agentic system (workflow planning, memory, non-linear task concatenation) to Dyna-2, coupled with full body control, with the goal of achieving ultra long duration full body autonomy rather than single step VLA For detailed information, please refer to https://www.dyna.co/dyna-2.1 It's still a very informative release