Diffusion Policy

Con­tin­u­ous Action Gen­er­a­tion

High-Dimensional ActionsContinuous actions are difficult to compose and reuse
Expert-Trajectory DependenceLong-tail task coverage is costly
Task FragmentationLimited transfer across scenes
Embodiment ChangesUsually require readaptation

KnowinGLOW

Autore­gres­sive Action Token

Unified RepresentationActions share a common token space
Native GenerationIntegrated autoregressive modeling
Compositional ReuseActions form a composable model language
Cross-Task TransferSupports task and embodiment scaling