Robion

“VLA models consist of two distinct stages: a Vision-Language Model (VLM) stage that is compute-intensive, and an Action Diffusion Transformer (ADiT) stage that is memory-intensive.”

“The authors profiled the 𝜋0 model and found that the VLM stage only achieves 52% average tensor core utilization, while the ADiT stage achieves only 15%”