Real-Time Chunking

A method that computes the next action chunk while still executing the current one, blended smoothly into what's already running.

Real-Time Chunking was proposed in 2025 by Kevin Black and colleagues at Physical Intelligence as an inference-time algorithm (NeurIPS 2025). Computing one action chunk with a large model takes on the order of a hundred milliseconds; executing chunks synchronously leaves a pause between them, while executing them asynchronously risks a jump where the new chunk doesn't match the old one. RTC starts computing the next chunk while the current one is still executing: the few steps that will inevitably run during inference are 'frozen,' and the rest is treated as an inpainting problem — a guidance term is added during flow-matching denoising, with a soft mask that gradually relaxes the constraint, so the new chunk blends smoothly with the old one. It needs no retraining; the paper demonstrated on π0.5 that the robot could still strike a match and light a candle with more than 300 ms of latency. A training-time version of RTC followed in December 2025, which simulates latency during training and removes the need for guidance computation at inference time.

ExampleTurning on the RTC setting for π0.5 in LeRobot, and setting the delay-step count to match measured inference time, means the robot no longer pauses between action chunks while executing them.

Also called

RTC, Real-Time Execution of Action Chunking Flow Policies

Related

Action Chunking · Asynchronous Inference · Inference Latency · Flow Matching · Temporal Ensembling · π0.5

Sources

Real-Time Execution of Action Chunking Flow Policies (arXiv 2506.07339)
Training-Time Action Conditioning for Efficient Real-Time Chunking (arXiv 2512.05964)
LeRobot Docs: Real-Time Chunking (RTC)

As of

2026-09