Dyna 的物理智能闭环：预测世界、控制身体，再接受真实客户的检验

A world-action model predicts future visual states as well as robot actions, using the predicted world to guide physical behavior.

Diffusion-based action generation is useful because robotic motion must be precise and smooth, while flow matching can simplify the generation process.

Human video provides broad physical priors, but a robot still needs its own demonstrations to bridge the gap between human movement and its embodiment.

Synthetic environments and self-generated video may expand training data, but their value depends on whether they improve real robotic tasks.

A large training model can be distilled into a smaller on-device model because physical control cannot rely on unpredictable network latency.

Dyna buys mature components with clear interfaces, but seeks to own hardware and inference decisions that determine the system's performance.

A wheeled semi-humanoid trades some human-like mobility for stability and practicality while preserving upper-body motions that support human-data transfer.

More video cannot supply missing tactile information, and overheating hands can slow both deployment and the experiments needed to improve a robot.

An agentic harness coordinates high-level decisions and waiting states so that low-level robotic skills can form a longer autonomous workflow.

A customer expanding a robot across multiple sites is stronger commercial evidence than a single pilot, although deployment claims still need independent verification.