Robot control needs both understanding the scene and reacting quickly. Understanding, like knowing where objects are, changes slowly. Reacting, like adjusting a grip, needs fast updates up to 50 times per second. Most robot models run their big understanding network every time they decide an action, which is too slow for reacting.
Leap-VLA splits these tasks. A large vision-language model runs slowly to understand the scene, and its output is reused. A fast expert uses this reused understanding along with fresh, quick video data to generate actions many times per second. The expert is trained to handle outdated understanding by knowing how old it is.
This method recovers most of the performance lost by using old scene information. On a standard robot benchmark, success rate jumped from 61.4% to 95.5% when the slow understanding refreshed every two seconds. This allows robots to react quickly without needing to run huge models each time, making them faster and more efficient.