Hierarchical Vision Language Action Models for Robotic Intelligence

Vision Language Action (VLA) Models utilize large pre-trained Vision Language backbones for robotic control. Due to inference-speed costs, hierarchical VLAs--a slow Vision Language Model conditioning a fast action expert--have emerged as a superior model class for high-frequency control. While performance benefits appear to stem from faster inference speed, we find that training dynamics are fundamentally altered and demonstrate how varying latency and speed at train time affects downstream performance. We perform interpretability-based probes to determine which components are most responsible for different model functions.