GeoAAC is a method for Vision-Language-Action (VLA) policies that use for action generation. Instead of a fixed action horizon, it adaptively determines how many predicted actions to execute before replanning. It does this by analyzing the geometry of the denoising trajectories, which reflects prediction reliability. The method requires no additional training and works from a single generation.
Experiments on GR00T N1.5 and π0.5 across simulation and real-world tasks show consistent gains: up to 8.7 percentage points in simulation, and real-world average success rising from 53.3% to 74.4% compared to fixed horizons and existing adaptive methods. This is interesting because it addresses the stage-dependent trade-off between action continuity and closed-loop responsiveness without extra supervision.