| """Real-Time Chunking (RTC) for action-chunking flow policies. |
| |
| Black, Galliker, Levine. "Real-Time Execution of Action Chunking Flow |
| Policies." NeurIPS 2025. arXiv:2506.07339. |
| |
| This package is the *backend-agnostic* RTC core plus an adapter for openpi's |
| PyTorch pi0.5. You bring your own pi0.5 base checkpoint (loaded through openpi); |
| RTC only replaces the inference path so consecutive action chunks stay |
| consistent while executing in real time. |
| |
| Layout |
| ------ |
| masking soft prefix masks + the PiGDM guidance weight (backend-agnostic) |
| controller Algorithm 1: async execution / background inference thread |
| sampler guided-inpainting denoising loop, PyTorch (torch.autograd.grad) |
| pi0 adapter: openpi ``PI0Pytorch`` pi0.5 (torch + openpi imported lazily) |
| |
| Importing this package pulls in neither torch nor openpi: the adapter and |
| sampler are imported lazily, so ``from rtc import RealTimeChunkingController`` |
| works in a plain-numpy process. |
| |
| See ``examples/pi05_realtime.py`` for an end-to-end, no-robot demo. |
| """ |
|
|
| __all__ = [ |
| "PrefixAttentionSchedule", |
| "RealTimeChunkingController", |
| "execution_horizon_for", |
| "get_prefix_weights_np", |
| "guidance_weight", |
| ] |
|
|
| from .controller import RealTimeChunkingController, execution_horizon_for |
| from .masking import PrefixAttentionSchedule, get_prefix_weights_np, guidance_weight |
|
|
