Rhoda AI: 1000X less training data required?

In this talk

1. 0:45

   Why Rhoda targets factories and warehouses

   Wooten says factories and warehouses offer stronger economics than home robotics, where customers may be unwilling to pay enough to support the hardware and service. These workplaces also have maintenance staff and face labor shortages; he cites 500,000 open manufacturing and logistics jobs in the United States. He favors bounded tasks as an early step toward more complex robot work.

2. 5:34

   A wheeled robot shaped by customer needs

   Rhoda says it studies customer environments and use cases to balance a useful design with flexibility. Its current concept is a wheeled humanoid, which Wooten says can be safer to stop, faster, more reliable, and able to carry more payload and battery capacity than a legged design. The company began with its AI model, then chose to develop hardware after evaluating more than 120 options.

3. 10:44

   Using internet video to teach physical skills

   Wooten describes robot learning as limited by the shortage of robot training data. Rhoda's approach uses internet video to build a model of physical motion: the robot predicts a video of a task, then converts that prediction into actions. He says the model runs at twice real time and can make updated predictions several times per second.

4. 25:35

   Industrial hardware and shorter training

   While its own hardware is under development, Rhoda plans to deploy its model on existing industrial equipment, including a certified seven-degree-of-freedom arm. Wooten identifies reliability, safety certification, payload, and reach as key hardware requirements. He says Rhoda has trained robots on complex tasks with 8 to 10 hours of robot data, compared with roughly 10,000 hours for some VLA examples, and aims to support tasks from text or visual instructions.