BLOG ILLUSTRATION / EXECUTION STORY

Predict a subtask and then generate a chunk

The VLA selects a language subtask. That subtask and the same current context condition motor generation. The shown inference uses fixed weights.

Predict a subtask and then generate a chunkOne VLA first selects a language subtask. The action expert then uses that subtask, the same current scene and the original instruction to generate continuous commands. The two context cards show reuse. They are not two sensors or a training update. 1 · Subtask selection 2 · Motor generation Current context Scene + instruction VLA prediction Select a subtask Language subtask pick up the pillow Same context Scene + instruction Action expert Continuous flow Action chunk Robot commands Subtask condition
The author made this π0.5 inference diagram from Physical Intelligence’s source figure. A language-level subtask explicitly conditions continuous action generation with the current observation and original instruction. The two context cards show reuse, not two sensors. Token-level details and official illustrative images are not shown. The playback is an example, and animation time is not measured runtime. Open full-size SVG