BLOG ILLUSTRATION / EXECUTION STORY

Training labels and runtime conditions

The diagram compares annotated training with two fixed-weight inference modes. Supported context may include an optional visual goal. That goal is not necessary for all actions.

Training labels and runtime conditionsThe training lane maps annotated examples to policy weights. Different inference modes use fixed weights and current context. A world-model visual goal is optional. Training annotations are not deployment inputs. Training does not generate an action chunk here. Training · learn the policy Training example Language + annotations π0.7 training Learn policy weights Policy weights Fixed during inference Goals · quality · speed Inference · use supported conditions Runtime context Scene + memory + intent World model Optional visual goal π0.7 policy Generate action chunk Fixed weights Keep current context
The author made this π0.7 conditioning diagram from Physical Intelligence’s source figure. Training uses labeled instructions, subgoal images and episode metadata. Inference can use task/subtask context, observation memory, desired metadata and world-model subgoals in a supported mode. The prompt interface receives these conditions before continuous action generation. Official performance demos are not shown, and visual subgoals are not necessary for all actions. The playback is an example, and animation time is not measured runtime. Open full-size SVG