Physical Intelligence Develops Real-Time Chunking Method with Inference-Time Freezing to Speed Up π0 and π0.5 VLAs in Robotics

Researchers at Physical Intelligence have developed a method called Real-Time Chunking (RTC) to address the challenge of high inference latency in vision-language-action (VLA) models used in robotics. These models typically experience delays and jerky transitions when processing actions, hindering smooth and real-time operation. RTC enables robots to perform inference and plan subsequent actions simultaneously while moving, effectively reducing delays and improving fluidity in action execution. The method uses an inference-time freezing and inpainting scheme to ensure smooth asynchronous action execution. This advancement applies to the π0 and π0.5 variants of VLA models, with RTC significantly speeding up the π0.5 model. Additionally, related research highlights a new multi-modal AI agent that employs a role-based workflow and Chain-of-LoRA strategy to efficiently analyze and reason over long videos, achieving better grounding accuracy than larger models.