Robots and self-driving cars use AI to create visual plans, like points or paths, to guide their actions. Generating these plans one coordinate at a time is slow and limits how often they can update their plans.
VFlash uses a lightweight "drafter" that predicts all coordinates of a plan in one step. It does not need to match the token order of a standard model.
This makes generation about 25 times faster for 3D point forecasts, and more accurate. Downstream tasks, like robot manipulation, succeed more often.