Two thresholds for robots: learning from human video and entering long shifts
Dyna Robotics reported results for a foundation model trained on human video, while Agility Robotics detailed battery and workplace claims for Digit V5 near human workers.
Artificial Intelligence··Evening
First-person human video instead of robot data
According to Interesting Engineering, Dyna Robotics trained its DYNA-2 robot foundation model on more than 1 million hours of first-person human video. The company compares that scale with roughly 170 years of continuous waking experience. The model predicts the next frame and the next action together in one world-modeling architecture, aiming to learn how people interact with objects from a broad video record. In Dyna's reported measurements, success on high-precision manufacturing tasks rose from 20 percent to a range of 80 percent to 90 percent through increased pre-training scale alone. In a zero-shot customer deployment, the quality pass rate was reported at 87 percent for DYNA-2, compared with 46 percent for DYNA-1, while co-training with video improved instruction-following scores by 133 percent. The company also says 13 minutes of data was enough to teach robotic hands to twist open a bottle cap. All of those figures come from the company, and the report includes no independent verification. Robots running the earlier DYNA-1 are already deployed in hotels, restaurants, and laundromats, placing the new model as a proposed next software layer for existing service settings rather than only a laboratory task.[1]
Digit V5 targets work approaching three shifts
Agility Robotics plans to ship Digit V5 to early customers in December. The company says its new battery system is designed for more than 20 hours of work per day, which would in principle let one unit cover three shifts. Upgraded cameras, sensors, and software are also said to let Digit work near people without the safety fencing commonly used around industrial robots. Agility reports that earlier Digit versions have logged more than 65,000 hours across nine sites. V5 uses the company's own cycloidal actuators in its hips and knees and carries an ISO-standard flange for fitting different end effectors. Automatic tool changes remain a plan for a later release. The report gives no payload figure, safety certification, or named customer site, and the operating claims come from the company. Agility also announced a special-purpose acquisition company deal valuing the company at 2.5 billion dollars. The product schedule, long-working-day claim, and new financing therefore arrived in the same announcement period. December shipments would mark the next reported step from accumulated site hours to the V5 hardware.[2]
Learning scale and workplace endurance are different tests
The two reports describe distinct thresholds as robots move closer to human environments. For DYNA-2, the input is a large pool of video recording what people see, rather than a robot's own action data. Its outcomes concern task success, instruction following, and a quality pass rate in a new customer deployment. For Digit V5, the announced threshold concerns how long the hardware can operate during a day, how closely it can be placed beside people, and how it can accept different end effectors. The first report addresses transfer from learned behavior; the second addresses entry of a physical system into shift work. There is no reported product link between them, and nothing says DYNA-2 runs Digit V5. Their common feature is the use of human activity as a reference point: one company turns human video into training input, while the other describes a robot's place within human shifts and shared work areas. The high performance rates, long operating time, and fence-free work claim all come from the companies. The picture presents two lines of progress in software learning and field hardware, both requiring independent testing.[1], [2]
