大晓联合香港大学发布StreamPI，让 VLA 真正理解时间，迈向连续物理智能

TL;DR - ACE Robotics and the University of Hong Kong introduced StreamPI, a streaming temporal-modeling method that gives vision-language-action models persistent multimodal context without adding model parameters. It improves robotic memory, spatial reasoning, and long-horizon task execution by reusing cached history instead of repeatedly processing frame windows.

• StreamPI anchors instructions to each observation, restructures attention across time, and uses key-value caching so inference processes only new observations while retaining historical representations.
• Random-interval streaming training exposes the model to variable observation gaps and incomplete histories, improving robustness to real-world timing and latency variations.
• On LIBERO, five-frame average success rose from 97.0% to 98.3%; on CALVIN, average completed task-chain length increased from 4.313 for π0.5 to 4.547.
• In real-robot tests, StreamPI improved shell-game success from 46.7% to 80.0%, rolling-object grasping from 26.7% to 63.3%, and cup-sleeve insertion from 60.0% to 92.0%.