Large robot policies produce chunks of future actions, but inference takes time. If the robot waits for each new chunk, motion stutters; if it switches chunks abruptly, it jerks.

Real-Time Chunking (Black, Galliker & Levine, Physical Intelligence, 2025) treats the next chunk as an inpainting problem: the actions already committed are kept fixed, and the model fills in the rest so the new chunk continues smoothly from the old one. The authors report smooth execution even with inference delays above 300 ms.

Why this matters: it is one concrete answer to a broader systems question — which decisions must happen on the robot in milliseconds, and which can wait for a slower, smarter model.